text
stringlengths 2
6.93k
| system_prompt
stringclasses 1
value |
|---|---|
2023. Llama 2: Open foundation and fine-tuned chat models.
# David Wadden, Shanchuan Lin, Kyle Lo, Lucy Lu Wang, Madeleine van Zuylen, Arman Cohan, and Hannaneh Hajishirzi. 2020. Fact or fiction: Verifying scientific claims. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7534–7550, Online. Association for Computational Linguistics.
# Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Majumder, and Furu Wei. 2022. Text embeddings by weakly-supervised contrastive pre-training. arXiv preprint arXiv:2212.03533.
| Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
AAAI Press.
|Shahul Es, Jithin James, Luis Espinosa-Anke, and Steven Schockaert|2023|Ragas: Automated evaluation of retrieval augmented generation|
|---|---|---|
|Angela Fan, Yacine Jernite, Ethan Perez, David Grangier, Jason Weston, and Michael Auli|2019|ELI5: Long form question answering|
|Adam Fisch, Alon Talmor, Danqi Chen, Eunsol Choi, Minjoon Seo, Patrick Lewis, Robin Jia, and Sewon Min, editors|2021|Proceedings of the 3rd Workshop on Machine Reading for Question Answering|
|Tianyu Gao, Howard Yen, Jiatong Yu, and Danqi Chen|2023|Enabling large language models to generate text with citations|
|Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, Qianyu Guo, Meng Wang, and Haofen Wang|2024|Retrieval-augmented generation for large language models: A survey|
|Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang|2020|Realm: Retrieval-augmented language model pre-training|
|Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov|2019|Natural questions: A benchmark for question answering research|
|---|---|---|
|Kenton Lee, Ming-Wei Chang, and Kristina Toutanova|2019|Latent retrieval for weakly supervised open domain question answering|
|Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela|2021|Retrieval-augmented generation for knowledge-intensive nlp tasks|
|Stephanie Lin, Jacob Hilton, and Owain Evans|2022|TruthfulQA: Measuring how models mimic human falsehoods|
|Yi Liu, Lianzhe Huang, Shicheng Li, Sishuo Chen, Hao Zhou, Fandong Meng, Jie Zhou, and Xu Sun|2023|Recall: A benchmark for llms robustness against external counterfactual knowledge|
|Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed|2023|Mistral 7b|
|---|---|---|
|Sayali Kulkarni, Sheide Chammas, Wan Zhu, Fei Sha, and Eugene Ie|2020|Aquamuse: Automatically generating datasets for query-based multi-document summarization|
|Chaitanya Malaviya, Subin Lee, Sihao Chen, Elizabeth Sieber, Mark Yatskar, and Dan Roth|2023|Expertqa: Expert-curated questions and attributed answers|
|Christopher D. Manning, Prabhakar Raghavan, and Hinrich Schütze|2008|Introduction to Information Retrieval. Cambridge University Press, Cambridge, UK|
|Sewon Min, Julian Michael, Hannaneh Hajishirzi, and Luke Zettlemoyer|2020|AmbigQA: Answering ambiguous open-domain questions| | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)
(EMNLP), pages 5783–5797, Online. Association for Computational Linguistics.
Fabio Petroni, Aleksandra Piktus, Angela Fan, Patrick Lewis, Majid Yazdani, Nicola De Cao, James Thorne, Yacine Jernite, Vladimir Karpukhin, Jean Maillard, Vassilis Plachouras, Tim Rocktäschel, and Sebastian Riedel. 2021. KILT: a benchmark for knowledge intensive language tasks. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2523–2544, Online. Association for Computational Linguistics.
Pranav Rajpurkar, Robin Jia, and Percy Liang. 2018. Know what you don’t know: Unanswerable questions for squad.
Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383–2392, Austin, Texas. Association for Computational Linguistics.
S. Robertson. 2009. The Probabilistic Relevance Framework: BM25 and Beyond. Foundations and Trends® in Information Retrieval, 3(4):333–389.
Anna Rogers, Matt Gardner, and Isabelle Augenstein. 2023. Qa dataset explosion: A taxonomy of nlp resources for question answering and reading comprehension. ACM Comput. Surv., 55(10).
Ivan Stelmakh, Yi Luan, Bhuwan Dhingra, and Ming-Wei Chang. 2022. ASQA: Factoid questions meet long-form answers. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 8273–8288, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.
Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca.
Nandan Thakur, Nils Reimers, Andreas Rücklé, Abhishek Srivastava, and Iryna Gurevych. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
2021. Beir: A heterogeneous benchmark for zero-shot evaluation of information retrieval models. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, volume 1. Curran.
| Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandez, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. 2023. Llama 2: Open foundation and fine-tuned chat models.
Ellen M. Voorhees and Donna K. Harman. 2005. TREC: Experiment and Evaluation in Information Retrieval (Digital Libraries and Electronic Publishing). The MIT Press.
Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Majumder, and Furu Wei. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
2024. Text embeddings by weakly-supervised contrastive pre-training.
Howard Yen, Tianyu Gao, Jinhyuk Lee, and Danqi Chen. 2023. MoQA: Benchmarking multi-type open-domain question answering. In Proceedings of the Third DialDoc Workshop on Document-grounded Dialogue and Conversational Question Answering, pages 8–29, Toronto, Canada. Association for Computational Linguistics. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# Question:
could never take the place of your man lyrics meaning
# Document Title:
Could Never Take the Place of Your Man
# Title used in answer:
Could Never Take the Place of Your Man
# Paragraph Sentences:
The song is an upbeat pop number featuring a combination of live drumming with two drum machine patterns. Also featured are two guitar solos - one wild and energetic, and one more bluesy and subdued in the full album cut. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
The song also features small elements of alternative music. The song consists of two verses and two choruses, followed by a lengthy instrumental coda. The lyrics paint the image of a woman seeking a man to replace the one who left, while Prince refuses, saying that she would not be satisfied with a one-night stand. The music and accompanying music video pushed this song to the top 10 in the US. The video was pulled from the Sign of the Times film and is a live take of the song and included the horn section of Eric Leeds and Atlanta Bliss.
# Type your answer here (it should be concise and only come from the passage/title):
The lyrics for the upbeat pop song "Could Never Take the Place of Your Man" paint the image of a woman seeking a man to replace the one who left, while Prince refuses, saying that she would not be satisfied with a one-night stand.
# How would you describe the question/answer? | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
(required)
- Complete Answer
# Figure 2: The Round 1 annotation task for CLAPNQ.
The annotator had to select the title/sentences needed to answer the question, and then provide a concise answer.
# Annotation Tasks
All annotation tasks were performed using Appen. They are described in Section 3 and 5 of the main paper. We provide screenshots and further instructions below.
# A.1 Dataset Creation
The CLAPNQ dataset was created in two rounds. A screenshot of round 1 is shown in Figure 2 and Figure 4. A small handful of the questions (1 in train, and 9 in dev) are high-quality annotations from the initial pilot rounds. These examples have several reference answers.
# A.2 Human Evaluation
The human evaluation was performed on a portion of the dev and test sets. Human eval on the GOLD generation task is shown in Figure 3. The RAG version had two additional questions regarding passage relevance as described in Section 5. We plan on releasing the human evaluation annotations as part of the dataset release. The general instructions to the annotator were as follows: In this task, you will review the same question and passage and, for each one, rate the quality of the answer to the question. On each page, you will see 3 different answers to the same question. Read the question and passage and answer how well you are confident in the question, passage, and know the correct answer. For each model answer, (given the same context and passage): The answer to the model is in red. Please make your judgments on this red answer span. Indicate if the answer is an “I don’t know” or if the answer is completely incoherent. For each model response, answer the following questions on a scale of 1-4: 1) DO NOT USE THE PASSAGE TO ANSWER THIS QUESTION: Does the response to the question look appropriate, useful, concise, and complete? 2) Is the response faithful to the passage? Evaluate each metric independently. Finally, also perform a head-to-head comparison of the model responses by answering the following question for every pair of answers: Which response do you prefer in terms of faithfulness, appropriateness, and naturalness?
# B Prompts
The Flan-T5 (Chung et al., 2022) prompt which was used for most models is: {title}: {passage} Please answer a question about this article. If the question is unanswerable, say “unanswerable”. user: {question}, answer: | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
Model A
one of the causes of the german hyperinflationary period that occurred after world war was Hyperinflation. By November 1922, the value of money in circulation had fallen from €300 million before World War I to €20 million. The Reichsbank responded by the unlimited printing of notes, thereby accelerating the devaluation of the mark. In his report to London, Lord Abernon wrote, "In the whole course of history, no dog has ever run after its own tail with the speed of the Reichsbank." Germany went through its worst inflation in December 1923; the exchange rate was 4,200,000,000,000 (42 * 10^12) Marks. In 1922, the highest denomination was 50,000 Marks. By 1923, the highest denomination was 100,000,000,000,000 (10^14) Marks. In 1923, the rate of inflation hit 3.25 x 10^9 percent per month (prices double every two days), exactly the same rate the Mark had in 1914. Beginning on 20 November 1923, 1,000,000,000,000 old Marks were exchanged for Rentenmark, and 4.2 Rentenmarks were worth 1 US dollar.
Answer: The Reichsbank responded by the unlimited printing of notes, thereby accelerating the devaluation of the mark. Germany went through its worst inflation in 1923.
Did the model indicate that it doesn't know the answer? eg: do not know the answer (required)
Is this answer completely irrelevant or extremely incoherent? (required)
yes
Is the response coherent and natural? (required)
Mostly No Mostly Yes
I feel like I'm talking to a robot Ifeel like I'm talking to a person
complete? (required)
DO NOT USE THE PASSAGE TO ANSWER THIS QUESTION: Does the response to the question look appropriate, useful, concise, and Mostly No Mostly Yes Yes not concise/complete Understood the question and responded concisely and appropriately
Is the response faithful to the passage? (required)
Mostly No Mostly Yes
Hallucinating Faithful E reference context passage
Optionally; please share any additional feedback regarding the answer and/or metrics:
Figure 3: The human evaluation task used to compare the model answers in random order. The individual questions per answer are shown here for one model.
The GPT Prompt is based on chat completion from that you do not have an answer.
OpenAI9: {'role': 'system', 'content': "Generate next agent response, given the following document(s). If you cannot base your answer on the document, please state that you do not have an answer.'}, {'role': 'system', 'content': "[title]: {title} [document]: {passage}, {'role': 'user', 'content': question}"}
The Llama Prompt is the default Llama 2 prompt (Touvron et al., 2023): <<SYS>> You are a helpful, respectful, and honest assistant. Always answer as helpfully as possible, while being safe. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
Your answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure that your responses are socially unbiased and positive in nature. If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information. <</SYS>> [document]: {title} {passage}. [conversation]: question},. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muennighoff. 2023. C-pack: Packaged resources to advance general Chinese embedding.
# Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W. Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. 2018. HotpotQA: A dataset for diverse, explainable multi-hop question answering. In Conference on Empirical Methods in Natural Language Processing (EMNLP).
# Peitian Zhang, Shitao Xiao, Zheng Liu, Zhicheng Dou, and Jian-Yun Nie. 2023. Retrieve anything to augment large language models.
# Ming Zhong, Yang Liu, Da Yin, Yuning Mao, Yizhu Jiao, Pengfei Liu, Chenguang Zhu, Heng Ji, and | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
Answer with no more than 150 words. If you cannot base your answer on the above document(s), please state
9https://learn.microsoft.com/en-us/azure/ ai-services/openai/reference 10https://huggingface.co/docs/transformers/model_ doc/flan-t5 | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
Document Title: Amazon Web Services
Paragraph Sentences:
- paid subscription
- AWS
- bonci
Original Answer:
paid subscription basis. The browser exits and flow into the virtual computer
I should complete
Update the answer here needed
OS netconn
Did you need to edit the answer (check all that are applicable):
- No
- Yes
- Yes added missing information
Figure 4: The Round 2 annotation task for CLAPNQ. The annotator had to verify and update the answer provided in Round 1 if needed. They also had to provide how they edited the answer.
The E5-CLAPNQ-T5-LG and E5-G-CLAPNQ-T5-LG models (<10%). We accommodate for these experiments with the longer context size by using a batch size of 8 and 10 epochs.
Examples
We provide several examples of output comparing the various LLMs as described in Section 6. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
See Figures 8-16. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# Question: one of the causes of the german hyperinflationary period that occurred after world war
Hyperinflation
By November 1922, the value thereby accelerating the devaluation of the mark. In his report to London gold of money in circulation had fallen from 300 million before World War I to €20 million. The Reichsbank responded by the unlimited printing note. Lord D'Abernon wrote In the whole course of history dog has ever run after its own tail with the speed of the Reichsbank. By 1923, the highest denomination was 100,000,000,000,000 (10) Marks. In December 1923 the exchange rate was Germany went through its worst inflation in 1923. In 1922, the highest denomination was 50,000 Marks. 2oooooOOO,OOO (4.2*10 Marks US dollar. In 1923, the rate of inflation hit 3.25 10 percent per month (prices double every two days exactly the same rate the Mark had in 1914. Beginning on 20 November 1923 1,000,0OO,0OO,OOO old Marks were exchanged for Rentenmark that 4.2 Rentenmarks were worth US dollar
# AB: Which response do you prefer in terms of faithfulness, appropriateness and naturalness? (required)
The Reichsbank responded by the unlimited printing of notes; thereby accelerating the devaluation of the mark Germany went through its worst inflation in 1923
The Reichsbank made unlimited printing of notes; thereby accelerating the devaluation of the mark - Then Germany went through its worst inflation in 1923.
No Preference (use sparingly)
# AC: Which response do you prefer in terms of faithfulness, appropriateness and naturalness? (required)
The Reichsbank responded by the unlimited printing of notes; thereby accelerating the devaluation of the mark Germany went through its worst inflation in 1923 the unlimited printing of notes by the Reichsbank This led to the rapid devaluation of the mark and contributed the severe hyperinflation experienced by Germany the early 1920s
No Preference (use sparingly)
# BC: Which response do you prefer in terms of faithfulness, appropriateness and naturalness? (required)
The Reichsbank made unlimited printing of notes; thereby accelerating the devaluation of the mark Then Germany went through its worst inflation in 1923. the unlimited printing of notes by the Reichsbank This led to the rapid devaluation of the mark and contributed the severe hyperinflation experienced by Germany the early 1920s.
Preference (use sparingly)
# Figure 5: The human evaluation task used to compare the model answers in random order. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
The head-to-head comparison for win-rate is shown here.
# Question: who are the characters in the black cat
Passage: Sephiria Arks (Sefiria Akusu), known as Number I, is the commander of the Numbers and is trusted enough by Chronos to be given total control over killing Creed. Sephiria was one of the few members who were "born for Chronos", as she stated that when she was an infant she underwent surgery to enhance her recovery time from injuries. Sven once commented that her eyes allowed her to stare into the recesses of one's soul. This empathy is her greatest strength, which allows her to utilize her powers of manipulation to their utmost extent. She will do anything to finish her missions and will sacrifice whatever it takes in order to succeed in meeting an objective. However, she does show some inner conflict in her manipulation, even questioning Belze about her actions. She controls all of the Chronos Numbers' actions as well as many of the independent operatives in the hunt for Creed with little interference from the Elders.
# Responses
GPT 3.5
In the Black Cat series, there are several characters. One of the main characters is Sephiria Arks, also known as Number I. She is the commander of the Numbers and has complete control over killing Creed. Sephiria underwent surgery as an infant to enhance her recovery time from injuries and has the ability to stare into the recesses of one's soul. She is highly skilled in manipulation and will do whatever it takes to succeed in her missions. Sephiria controls the actions of the Chronos Numbers and many independent operatives in the hunt for Creed.
CLAPNQ-T5-LG
Sephiria Arks, known as Number I, is the commander of the Numbers and is trusted enough by Chronos to be given total control over killing Creed. Sven once commented that her eyes allowed her to stare into the recesses of one's soul. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
This empathy is her greatest strength, which allows her to utilize her powers of manipulation to their utmost extent.
Reference
No answer
# Table 8: This is an Unanswerable question that is actually partially answerable from the passage. It only mentions one character. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# Question: what type of volcano is one tree hill
Passage: Maungakiekie One Tree Hill Maungakiekie / One Tree Hill after the removal of the “one tree”. Highest point Elevation 182 m (597 ft) Coordinates 36°54’0” S 174°46’59” E · / · 36.90000°S 174.78306°E · / - 36.90000; 174.78306 Coordinates: 36°54’0” S 174°46’59” E · / · 36.90000°S 174.78306°E · / - 36.90000; 174.78306 Geography Location North Island, New Zealand Geology Volcanic arc/belt Auckland volcanic field Location
# Responses
|GPT 3.5|One Tree Hill is part of the Auckland volcanic field, which is a volcanic arc/belt.|
|---|---|
|CLAPNQ-T5-LG|No answer|
|Reference|No answer|
Table 9: This is a difficult example which confused most annotators during the human evaluation. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
It appears from the passage that the GPT answer may be correct but the answer is “scoria cone volcano” which is not in the passage. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
Question: who is given credit for inventing the printing press
Passage: The world’s first movable type printing press technology for printing paper books was made of porcelain materials and was invented around AD 1040 in China during the Northern Song Dynasty by the inventor Bi Sheng (990 – 1051). Subsequently in 1377, the world’s oldest extant movable metal print book, Jikji, was printed in Korea during the Goryeo dynasty. Because of this, the diffusion of both movable-type systems was, to some degree, limited to primarily East Asia, although various sporadic reports of movable type technology were brought back to Europe by Christian missionaries, traders and business people who were returning to Europe after having worked in China for several years and influenced the development of printing technology in Europe. Some of these medieval European accounts are still preserved in the library archives of the Vatican and Oxford University among many others. Around 1450 Johannes Gutenberg, introduced the metal movable-type printing press in Europe, along with innovations in casting the type based on a matrix and hand mould. The small number of alphabetic characters needed for European languages was an important factor. Gutenberg was the first to create his type pieces from an alloy of lead, tin, and antimony – and these materials remained standard for 550 years.
Responses
GPT 3.5
The invention of pe printing press is commonly credited to Johannes Gutenberg, who introduced pe metal movable-type printing press in Europe around 1450. Gutenberg’s innovations in casting pe type based on a matrix and hand mould, along wip pe use of an alloy of lead, tin, and antimony for creating pe type pieces, are considered significant advancements in printing technology.
Mistral
Johannes Gutenberg of Mainz, Germany is given credit for inventing pe metal movable-type printing system in Europe and pe printing press.
CLAPNQ-T5-LG
The world’s first movable type printing press technology for printing paper books was invented around AD 1040 in China during pe Norpern Song Dynasty by pe inventor Bi Sheng (990 – 1051). Around 1450 Johannes Gutenberg introduced pe metal movable-type printing press in Europe.
Reference
The world’s first movable type printing press technology for printing paper books was made of porcelain materials and was invented around AD 1040 in China during pe Norpern Song Dynasty by pe inventor Bi Sheng (990 – 1051). Around 1450 Johannes Gutenberg, introduced pe metal movable-type printing press in Europe.
Table 10: The GPT and Mistral answers look good but are only partially faithful to the passage, because they do not give any credit to the innovation in China that is also mentioned in the passage. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
The GPT answer is also too long. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# Claim Generation Prompting
A "claim" is a statement or assertion made within a text expressing a belief, opinion, or fact. Given evidence from the original context, please extract one claim and its associated topics.
| Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
Question: who played scotty in the new star trek movie
Passage: Urban was born in Wellington, New Zealand. His father, a German immigrant, owned a leather goods store, and his mother once worked for Film Facilities in Wellington. Through his mother, the young Urban was exposed to classic New Zealand cinema, and developed an interest in the film industry. Urban attended St Mark’s Church School, where he showed an early love for public performance. His first acting role came at age eight, when he had a single line in one episode of the New Zealand television series Pioneer Woman. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
Though continuing to take part in school stage productions, he did not act professionally again until after high school.
|Responses|
|---|
|GPT 3.5|Simon Pegg played Scotty in the new Star Trek movies.|
|CLAPNQ-T5-LG|No answer|
|Reference|No answer|
Table 11: GPT finds the correct answer but it is not grounded in the document. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# Question: where would a commercial quantity of cobalt-60 be produced
Passage: Cobalt - 60 , Co , is a synthetic radioactive isotope of cobalt with a half - life of 5.2714 years . It is produced artificially in nuclear reactors . | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
Deliberate industrial production depends on neutron activation of bulk samples of the monoisotopic and mononuclidic cobalt isotope Co . Measurable quantities are also produced as a by - product of typical nuclear power plant operation and may be detected externally when leaks occur . In the latter case ( in the absence of added cobalt ) the incidentally produced Co is largely the result of multiple stages of neutron activation of iron isotopes in the reactor ’s steel structures via the creation of Co precursor . The simplest case of the latter would result from the activation of Fe . Co decays by beta decay to the stable isotope nickel - 60 ( Ni ) . The activated nickel nucleus emits two gamma rays with energies of 1.17 and 1.33 MeV , hence the overall nuclear equation of the reaction is Co +n → Co → Ni +e + ν+ gamma rays .
# Responses
|GPT 3.5|A commercial quantity of cobalt-60 is typically produced through neutron activation of bulk samples of the monoisotopic and mononuclidic cobalt isotope Co. This industrial production process usually takes place in nuclear reactors. Additionally, measurable quantities of cobalt-60 can also be produced as a by-product of typical nuclear power plant operations. However, it’s important to note that cobalt-60 is highly regulated and controlled due to its radioactive nature.|
|---|---|
|CLAPNQ-T5-LG|Cobalt-60 is produced artificially in nuclear reactors. Measurable quantities are also produced as a by - product of typical nuclear power plant operation.|
|Mistral|Cobalt-60 is typically produced in nuclear reactors through the irradiation of natural cobalt. Commercial quantities are produced by specialized companies in several countries including Argentina, Canada, and Russia, which are the largest suppliers of Cobalt-60 in the world. These companies have dedicated facilities for the production of Cobalt-60 using nuclear reactors. The irradiated cobalt is then extracted and purified to produce the commercial product.|
|Reference|It is produced artificially in nuclear reactors. Measurable quantities are also produced as a by - product of typical nuclear power plant operation.| | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# Question: where are nimbus clouds found in the sky
# Passage 1 (GOLD):
Nimbostratus has a diffuse cloud base generally found anywhere from near surface in the low levels to about 3,000 m (9,800 ft) in the middle level of the troposphere. Although usually dark at its base, it often appears illuminated from within to a surface observer. Nimbostratus usually has a thickness of about 2000 to 4000 m. Though found worldwide, nimbostratus occurs more commonly in the middle latitudes. It is coded C2 on the SYNOP report.
# Passage 2:
Nimbostratus occurs along a warm front or occluded front where the slowly rising warm air mass creates nimbostratus along with shallower stratus clouds producing less rain, these clouds being preceded by higher-level clouds such as cirrostratus and altostratus. Often, when an altostratus cloud thickens and descends into lower altitudes, it will become nimbostratus.
# Passage 3:
Multi-level nimbostratus is physically related to other stratiform genus-types by way of being non-convective in nature. However, the other sheet-like clouds usually each occupy only one or two levels at the same time. Stratus clouds are low-level and form from near ground level to 2,000 metres (6,600 ft) at all latitudes. In the middle level are the altostratus clouds that form from 2,000 metres (6,600 ft) to 7,000 metres (23,000 ft) in polar areas, 7,000 metres (23,000 ft) in temperate areas, and 7,600 metres (24,900 ft) in tropical areas...
# Responses
|GPT 3.5|Gold Nimbostratus clouds are generally found anywhere from near the surface in the low levels to about 3,000 m (9,800 ft) in the middle level of the troposphere. They can be found worldwide, but they occur more commonly in the middle latitudes.|
|---|---|
|RAG|Nimbostratus clouds are found in the middle level of the troposphere, typically anywhere from near the surface to about 3,000 meters (9,800 feet) in altitude.|
|Mistral|Gold Nimbostratus clouds are found in the middle levels of the troposphere, at altitudes ranging from near the surface up to about 3,000 meters (9,800 ft).|
| |Gold|RAG|
|---|---|---|
|CLAPNQ-T5-LG|Nimbostratus has a diffuse cloud base generally found anywhere from near surface in the low levels to about 3,000 m (9,800 ft) in the middle level of the troposphere. Though found worldwide, nimbostratus occurs more commonly in the middle latitudes.|Nimbostratus has a diffuse cloud base generally found anywhere from near surface in the low levels to about 3,000 m (9,800 ft) in the middle level of the troposphere. Stratus clouds are low-level and form from near ground level to 2,000 metres (6,600 ft) at all latitudes. In the middle level are the altostratus clouds that form from 2,000 metres (6,600 ft) to 7,000 metres (23,000 ft) in polar areas, 7,000 metres (23,000 ft) in temperate areas, and 7,600 metres (24,900 ft) in tropical areas.|
# Reference
Nimbostratus has a diffuse cloud base generally found anywhere from near surface in the low levels to about 3,000 m in the middle level of the troposphere. Though found worldwide, nimbostratus occurs more commonly in the middle latitudes.
# Table 13:
In this example we compare the responses when just the GOLD answer is provided and the top 3 passages are provided when the first passage is the GOLD passage. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
All models provide good responses when only the gold answer is provided. In the non-RAG setting CLAPNQ-T5-LG and Mistral provided irrelevant information from the other passages about other clouds. All three missed the sentence that they can be found world-wide in the RAG setup. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# Question: who was glumdalclitch how did she help gulliver
GPT 3.5 Gold
Glumdalclitch was a nine-year-old girl, pe daughter of a giant farmer who captured Gulliver. She became fond of Gulliver and treated him like a pet, creating a miniature bed-chamber for him and playing wip him as if he were a doll. Glumdalclitch was also a skilled seamstress and made dolls’ clopes for Gulliver. While she provided companionship and amusement for Gulliver, it doesn’t mention any specific ways in which she helped him beyond pat.
Mistral Gold
Glumdalclitch was a nine-year-old girl from Brobdingnag who captured Gulliver and made him her pet. She was only forty feet tall, making Gulliver appear gigantic in comparison. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
Despite her small size, she was skilled in sewing and made clopes for Gulliver from pe finest Brobdingnagian fabric. However, pe fabric irritated Gulliver’s skin. Glumdalclitch was amused by Gulliver and enjoyed playing wip him as if he were a doll. She gave him pe pet name of Glumdalclitch, which means "little nurse" in pe Brobdingnagian language. Despite being a captive, Gulliver grew very fond of Glumdalclitch and was proud to perform stunts for her entertainment.
CLAPNQ-T5-LG Gold
Glumdalclitch is a skilled seamstress wip a talent for making dolls’ clopes. She makes Gulliver her pet, creates a small travel case for him, and is amused to play wip him as if he were a doll. Gulliver grows very proud of pe stunts pat he performs for Glumdalclitch’s amusement.
Reference: Glumdalclitch is a nine-year-old, forty feet tall daughter of the farmer who captured Gulliver. She made Gulliver her pet, created a bed for him out of a travel case and made him clothes.
Table 14: An example that can be used with Table 15 to compare the GOLD setup with the RAG setup when the GOLD passage is not retrieved during RAG. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# Question: who was glumdalclitch how did she help gulliver
Passage 1: When the Queen of Brobdingnag takes Gulliver into her court, he has Glumdalclitch brought to court with him. The prideful Gulliver thinks of himself as being greatly honored and promoted by moving to court, but never ceases to love and seek the approval of the little girl who first helped him. Indeed, he remembers her fondly even after returning to England.
Passage 2: Glumdalclitch is the name Gulliver gives his “nurse” in Book II of Jonathan Swift’s Gulliver’s Travels. In Book I, Gulliver travels to the land of Lilliput. Leaving there, he travels to the land of Brobdingnag. In Lilliput, Gulliver was a giant, and in Brobdingnag, he is a dwarf, with the proportions reversed.
Passage 3: This article is written like a personal reflection or opinion essay that states a Wikipedia editor’s personal feelings about a topic. Please help improve it by rewriting it in an encyclopedic style. (October 2017) (Learn how and when to remove this template message) Gulliver
# Responses
|GPT 3.5 RAG|Glumdalclitch was a character in Jonathan Swift’s novel Gulliver’s Travels. She was the nurse or caretaker of Gulliver in the land of Brobdingnag. Glumdalclitch helped Gulliver by taking care of him, ensuring his safety, and assisting him during his time in the court of the Queen of Brobdingnag. While Gulliver felt honored and promoted in the court, he still cherished and sought approval from Glumdalclitch, showing his deep affection for her.|
|---|---|
|Mistral RAG|Glumdalclitch was a little girl whom Gulliver met in Brobdingnag, a land of giants, where he was a dwarf. She helped him by nursing him and taking care of him when he was ill. After Gulliver’s return to England, he remembered her fondly.|
|CLAPNQ-T5-LG RAG|Glumdalclitch is the name Gulliver gives his nurse” in Book II of Jonathan Swift’s Gulliver’s Travels. In Lilliput, Gulliver was a giant, and in Brobdingnag, he is a dwarf, with the proportions reversed.|
Reference: Glumdalclitch is a nine-year-old, forty feet tall daughter of the farmer who captured Gulliver. She made Gulliver her pet, created a bed for him out of a travel case and made him clothes.
| Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# Table 15: Alternative appropriate and relevant answers based on the retrieved passages compared to the GOLD setup in Table 14. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
Note: The claim should not contain ambiguous references, such as 'he', 'she,' and 'it', and should use complete names. If there are multiple topics, give the most dominant one. The target of the claim (one entity) is the specific individual, group, or organization that the statement or assertion within a text is directed towards or about which it is making a case. The topic of the claim should be a simple phrase representing the claim’s central argument concept. If there is no claim, please leave it blank. Please generate a claim based on the given evidence. Don’t generate the evidence yourself.
| Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# Conversation
# User: why did the us demand trade with japan
# Passages
Passage 1
The United States reacted by seeking to bring the Japanese war effort to a complete halt by imposing a full embargo on all trade between the United States to Japan on 1 August 1941, demanding that Japan withdraw all troops from both China and Indochina. Japan was dependent on the United States for 80 percent of its oil, resulting in an economic and military crisis for Japan that could not continue its war effort with China without access to petroleum and oil products. Attack
Passage 2
The U.S. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
embargoes gave Japan a sense of urgency. It would either have to agree to Washington’s demands or use force to gain access to the resources it needed.
Passage 3
Japan’s goal after 1931 was economic dominance of most of East Asia, often expressed in Pan-Asian terms of “Asia for the Asians.” Japan was determined to dominate the China market, which the U.S. and other European powers had been dominating. On October 19, 1939, the American Ambassador to Japan, Joseph C. Grew, in a formal address to the America-Japan Society stated:
# Responses
gpt3.5-turbo
The United States demanded a halt to trade with Japan in response to Japan’s aggressive actions in East Asia and its invasion of China. The U.S. sought to put pressure on Japan to withdraw its troops from China and Indochina by imposing an embargo on trade. Additionally, the U.S. was concerned about Japan’s growing economic dominance in the region, particularly in the China market, which had previously been dominated by the U.S. and other European powers. By cutting off trade, the U.S. hoped to weaken Japan’s war effort and force a resolution to the conflict through diplomatic means.
mistral
The United States demanded trade with Japan not because it wanted to help Japan, but because it wanted to ensure its own economic interests and prevent Japan from dominating the China market and expanding its military power in the Pacific. The U.S. embargo on oil and other resources was a response to Japan’s aggression in China and its refusal to withdraw its troops from that region. This economic pressure forced Japan into a corner and ultimately led to its decision to attack the U.S. naval base at Pearl Harbor on December 7, 1941. The attack was a desperate move by Japan to secure the resources it needed to continue its war efforts.
flan-t5-large
Japan was dependent on the United States for 80 percent of its oil, resulting in an economic and military crisis for Japan that could not continue its war effort with China without access to petroleum and oil products. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# arXiv:2406.13213v1 [cs.CL] 19 Jun 2024
Multi-Meta-RAG: Improving RAG for Multi-Hop Queries using Database Filtering with LLM-Extracted Metadata
Mykhailo Poliakov[0009−0006−5263−762X] and Nadiya Shvai[0000−0001−8194−6196]
National University of Kyiv-Mohyla Academy
{mykhailo.poliakov, n.shvay}@ukma.edu.ua
Abstract. The retrieval-augmented generation (RAG) enables retrieval of relevant information from an external knowledge source and allows large language models (LLMs) to answer queries over previously unseen document collections. However, it was demonstrated that traditional RAG applications perform poorly in answering multi-hop questions, which require retrieving and reasoning over multiple elements of supporting evidence. We introduce a new method called Multi-Meta-RAG, which uses database filtering with LLM-extracted metadata to improve the RAG selection of the relevant documents from various sources, relevant to the question. While database filtering is specific to a set of questions from a particular domain and format, we found out that Multi-Meta-RAG greatly improves the results on the MultiHop-RAG benchmark. The code is available on GitHub.
Keywords: large language models · retrieval augmented generation · multi hop question answering
# 1 Introduction
Large Language Models (LLMs) have shown remarkable language understanding and generation abilities [8,11]. However, there are two main challenges: static knowledge [6] and generative hallucination [3]. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
Retrieval-augmented generation [4] is an established process for answering user questions over entire datasets. RAG also helps mitigate generative hallucination and provides LLM with a source of new information on which it was not trained [9]. Real-world RAG pipelines often need to retrieve evidence from multiple documents simultaneously, a procedure known as multi-hop querying. Nevertheless, existing RAG applications face challenges in answering multi-hop queries, requiring retrieval and reasoning over numerous pieces of evidence [10]. In this paper, we present Multi-Meta-RAG: an improved RAG using a database filtering approach with LLM-extracted metadata that significantly improves the results on the MultiHop-RAG benchmark. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# M. Poliakov and N. Shvai
# Related works
MultiHop-RAG [10] is a novel benchmarking dataset focused on multi-hop queries, including a knowledge base, questions, ground-truth responses, and supporting evidence. The news articles were selected from September 26, 2023, to December 26, 2023, extending beyond the knowledge cuto of ChatGPT1 and GPT-42. A trained language model extracted factual or opinion sentences from each news article. These factual sentences act as evidence for multi-hop queries. The selection method involves keeping articles with evidence that overlaps keywords with other articles, enabling the creation of multi-hop queries with answers drawn from numerous sources. Given the original evidence and its context, GPT-4 was used to rephrase the evidence, referred to as claims. Afterward, the bridge entity or topic is used to generate multi-hop queries.
For example, "Did Engadget report a discount on the 13.6-inch MacBook Air before The Verge reported a discount on Samsung Galaxy Buds 2?" is a typical query from the MultiHop-RAG dataset. Answering it requires evidence from Engadget and The Verge to formulate an answer. Also, it requires LLM to figure out the temporal ordering of events. MultiHop-RAG also has inference, comparison, and null (without correct answer) queries, in addition to the temporal query above.
Engadget Chunk
Did Engadget report a discount on pe 13.6-inch MacBook Air before The Verge reported a discount on Samsung Galaxy Buds 2?
|User|Prompt|
|---|---|
| |Query|
|News Article|News Article|
|---|---|
|"source": "BBC"|"source": "CNN"|
|News Article|News Article|
|---|---|
|"source": "The Verge"|LLM Wrong response|
Fig. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
1. A naive RAG implementation for MultiHop-RAG queries. RAG selects chunks from articles not asked in the example query, which leads to LLM giving a wrong response.
1 gpt-3.5-turbo-0613
2 gpt4-0613 | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# Multi-Meta-RAG
In a typical RAG application, we use an external corpus that comprises multiple documents and serves as the knowledge base. Each document within this corpus is segmented into chunks. These chunks are then converted into vector representations using an embedding model and stored in a vector database. Given a user query, RAG typically retrieves the top-K chunks that best match the query. The retrieved chunks, combined with the query are submitted into an LLM to generate a final response.
For the MultiHop-RAG benchmark, scraped articles act as a knowledge base for the RAG application tested. The problem is that a naive RAG application fails to recognize that the query asks for information from specific sources. Top-K chunks such as RAG retrieves often contain information from sources other than those mentioned in the query. Retrieved chunks might even miss relevant sources, leading to a wrong response, as depicted in Figure 1.
# Multi-Meta-RAG
Did Engadget report a discount on the 13.6-inch MacBook Air before The Verge reported a discount on Samsung Galaxy Buds 2?
|LLM|Extract Metadata|
|---|---|
|{ "source":{ "$in":[ "Engadget", "The Verge" ] } } |{ "source":{ "$in":[ "Engadget", "The Verge" ] } } |
|Embedding|
|---|
|News Article "source": "Engadget"|News Article "source": "BBC"|
|News Article "source": "The Verge"|News Article "source": "CNN"|
Store vectorized chunks and metadata in Vector DB (Neo4j) for Engadget, The Verge, BBC, CNN, etc.
Metadata
Engadget Chunk
The Verge Chunk
Correct response LLM. The Verge Chunk. The Verge Chunk.
Fig. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
2. Multi-Meta-RAG: an improved RAG with database filtering using metadata. Metadata is extracted via secondary LLM. With filtering, we can ensure top-K chunks are always from relevant sources with better chances of getting correct overall responses. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# M. Poliakov and N. Shvai
# Extraction of Relevant Query Metadata with the LLM
|Query|Extracted Filter|
|---|---|
|Does the TechCrunch article report on new hiring at Starz, while the Engadget article discusses layoffs within the entire video game industry?|"source": { " $in " : [ "TechCrunch", "Engadget" ] }|
|Did The Guardian’s report on December 12, 2023, contradict the Sporting News report regarding the performance and future outlook of Manchester United?|"published_at": { " $in " : [ "December 12, 2023" ] }, "source": { " $in " : [ "The Guardian", "Sporting News" ] }|
|Who is the individual facing a criminal trial on seven counts of fraud and conspiracy, previously likened to a financial icon but not by TechCrunch, and is accused by the prosecution of committing fraud for wealth, power, and influence?|"source": { " $nin " : [ "TechCrunch" ] }|
Each question in the MultiHop-RAG [10] benchmark follows a typical structure. Every query requests information from one or more sources of news. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
In addition, some temporal queries require news articles from a particular date. We can extract the query filter via helper LLM by constructing a few-shot prompt [1] with examples of extracted article sources and publishing dates as a filter. The prompt template is provided in Appendix A. We only run metadata extraction with ChatGPT3 because this additional RAG pipeline step must be quick and cheap. We found out that this step takes 0.7 seconds on average for one query.
Two query metadata filter fields are extracted: article source and publication date. The complete filter is a dictionary with two fields combined. Samples of extracted metadata filters can be found in Table 1. The primary filtering operator is $in, the only operator provided in the examples in a few-shot prompt template. The LLM also correctly chooses a tiny fraction of the $nin operator for some queries without an example. While LLM only used $in and $nin for article sources, the model sometimes chooses other operators like $lt or $gt for GPT-3.5-Turbo-1106. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
Please give the response following this format:
Evidence: [original context]
Claims: [extract claim]
Claim Target: [target]
Claim Topic: [topic]
Here are examples:
Now, it’s your turn.
News
evidence
# Inference Query Generation Prompting
A multi-hop question is a query requiring multiple inferential leaps or accessing several pieces of information from different locations or sources to arrive at an answer. The following are news articles’ metadata and claims come from the articles. All the claims from the article are related to a similar target. Your task is to generate one multi-hop inference question based on the claims. Here are some instructions:
1. Find the Connection: The connection between claims is <target>, which is how these key pieces of information are related or how they can be combined to form a more complex idea.
2. Formulate the Question: Create a question that cannot be answered by relying on just one of the sentences but instead requires understanding and linking the information from all of the sources. The answer is <target>.
3. Ensure Coherence: Make sure the question flows logically from the combined information and is clear and unambiguous.
4. Use the keywords: <key set>
Examples:
Context:
Context | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# Multi-Meta-RAG 5
Publication date for a fraction of temporal queries. Because the number of such queries is small, we decided to only use date filters with $in and $nin operators and a most frequent date format for easier matching in the database. All queries have a source filter extracted, while the publishing date filter was extracted in 15.57% of queries, while 22.81% of queries of the MultiHop-RAG dataset are temporal.
# Improved Chunk Selection using Metadata Filtering
The extracted metadata could be used to enhance an RAG application (Figure 2). We split the articles in the MultiHop-RAG knowledge base into chunks, each containing 256 tokens using LLamaIndex using a sentence splitter as in the original MultiHop-RAG implementation. We also picked a chunk overlap of 32, finding out that smaller chunk overlap leads to a better variety of unique chunks in the top-K selection than the original implementation, which used the LLamaIndex default of 200. We selected LangChain Neo4j vector store as a vector database as its index implementation recently started to support metadata filtering. We then convert the chunks using an embedding model and save the embeddings into a vector database with article metadata saved as node properties.
Likewise, in the retrieval stage, we transform a query using the same embedding model and retrieve the top-K most relevant chunks with the highest cosine similarity with the query embedding. We also filter the chunks with LLM-extracted metadata in the same stage. Similarly to MultiHop-RAG, we use a Reranker module (bge-reranker-large) to examine the retrieval performance. After retrieving 20 corresponding chunks using the embedding model and metadata filter, we select the top-K chunks using the Reranker.
# Results
# Chunk Retrieval Experiment
We selected two best-performing embedding models from the original MultiHop-RAG experiment for testing metadata filtering chunk retrieval performance, bge-large-en-v1.5 and voyage-02. The retrieved list of chunks is compared with the ground truth evidence associated with each query, excluding the null queries, as they lack corresponding evidence. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
For evaluation, we assume the Top-K chunks are retrieved and use metrics such as Mean Average Precision at K (MAP@K), Mean Reciprocal Rank at K (MRR@K), and Hit Rate at K (Hit@K). MAP@K measures the average precision of the top-K retrieval across all queries. MRR@K calculates the average of the reciprocal ranks of the first relevant chunk within the top-K retrieved set for each query. Hit@K measures the proportion of evidence that appears in the top-K retrieved set. The experiment (Table 2)
4 strftime format code %B %-d, %Y
5 April 2024 | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# M. Poliakov and N. Shvai
with RAG showed considerable improvement in both embeddings for all core metrics MRR@10, MAP@10, Hits@10, and Hits@4. Most notably, for voyage-02 Hits@4 enhanced by 18%. This is important for practical RAG systems, where the top-K retrieved should as as low as possible to account for context window limits and cost.
**Table 2. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
Chunk retrieval experiment results. Top-10 chunks are selected with bge-reranker-large after Top-20 chunks are found via the similarity search and database metadata filtering. A chunk size of 256 and a chunk overlap of 32 is used.**
|Embedding|Baseline RAG [10]|MRR@10|MAP@10|Hits@10|Hits@4|
|---|---|---|---|---|---|
|bge-large-en-v1.5 (reported)| |0.563|0.4759|0.7183|0.6364|
|voyage-02 (reported)| |0.586|0.4795|0.7467|0.6625|
|voyage-02 (repository sample)| |0.6152|0.2718 a|0.7315|0.6683|
**Embedding Multi-Meta-RAG (ours)**
| |MRR@10|MAP@10|Hits@10|Hits@4|
|---|---|---|---|---|
|bge-large-en-v1.5|0.6574|0.3293|0.8909|0.7672|
|voyage-02|0.6748|0.3388|0.9042|0.792|
Note a. We found a difference between reported MAP@10 and evaluated MAP@10 on the baseline voyage-02 retrieval sample file provided in the MultiHop-RAG repository. We take the evaluated value of 0.2718 for comparisons between RAGs.
# LLM Response Generation Experiment
|LLM|Accuracy|
|---|---|
|Ground-truth [10]| |Baseline RAG [10]| |Multi-Meta-RAG (ours)|
|GPT4 (gpt-4-0613)|0.89|0.56|0.63b c|
|PaLM (text-bison@001)|0.74|0.47|0.61b|
As with embeddings, we picked two best-achieving LLMs on ground-truth chunks based on MultiHop-RAG initial experiments, GPT-4 and Google PaLM. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# Multi-Meta-RAG 7
achieved substantial improvement in accuracy (Table 3) for both models compared to baseline RAG implementation. Google PaLM accuracy improved by 26% from 0.47 to 0.61. Preliminary GPT-4 results also show a 12% increase from 0.56 to 0.63.
Note b. The accuracy measuring script was not provided in the MultHop-RAG repository. We were able to recreate our script, which shows similar results for ground-truth chunks for both models and on experiments with the sample voyage-02 retrieval list provided in the repository. We accept the answer as correct when the model answer appears in the gold answer or the gold answer appears in the model answer. Both answers are transformed to lowercase, striped from leading and trailing whitespace, and have their punctuation removed before comparison. This script is used for benchmarking Multi-Meta-RAG.
Note c. GPT-4 accuracy is a preliminary result based on 50 queries (2% of all queries). Results on the complete dataset are pending.
# Conclusion
In this paper, we introduce Multi-Meta-RAG, a method of improving RAG for multi-hop queries using database filtering with LLM-extracted metadata. Multi-Meta-RAG considerably improves results in both chunk retrieval and LLM generation experiments, while also being quite straightforward and explainable. The proposed solution still has some limitations. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
Firstly, extracting metadata requires a set of queries from a particular domain and question format, as well as additional inference time. Secondly, it requires the manual creation of a prompt template that will extract the metadata from the query. Thirdly, while the improved results are encouraging, they still fall considerably below the results achieved by feeding LLM precise ground-truth facts. Future work may include testing more generic prompt templates for metadata extraction on a variety of multi-hop datasets from different domains. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# References
|1.|Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
33, pp. 1877–1901. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
Curran Associates, Inc. (2020)|
|---|---|
|2.|Chase, H.: LangChain (Oct 2022), https://github.com/langchain-ai/langchain|
|3.|Huang, L., Yu, W., Ma, W., Zhong, W., Feng, Z., Wang, H., Chen, Q., Peng, W., Feng, X., Qin, B., Liu, T.: A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions (2023)|
|4.|Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W.t., Rocktäschel, T., Riedel, S., Kiela, D.: Retrieval-augmented generation for knowledge-intensive nlp tasks. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 9459–9474. Curran Associates, Inc. (2020)|
|5.|Liu, J.: LlamaIndex (Nov 2022). https://doi.org/10.5281/zenodo.1234, https://github.com/jerryjliu/llama{_}index|
|6.|Mialon, G., Dessì, R., Lomeli, M., Nalmpantis, C., Pasunuru, R., Raileanu, R., Rozière, B., Schick, T., Dwivedi-Yu, J., Celikyilmaz, A., Grave, E., LeCun, Y., Scialom, T.: Augmented language models: a survey (2023)|
|7.|Neo4j, Inc.: Neo4j graph database, https://neo4j.com/product/neo4j-graph-database|
|8.|Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P.F., Leike, J., Lowe, R.: Training language models to follow instructions with human feedback. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems. vol. 35, pp. 27730–27744. Curran Associates, Inc. (2022)|
|9.|Shuster, K., Po , S., Chen, M., Kiela, D., Weston, J.: Retrieval augmentation reduces hallucination in conversation. In: Moens, M., Huang, X., Specia, L., Yih, S.W. (eds.) Findings of the Association for Computational Linguistics: EMNLP 2021, Virtual Event / Punta Cana, Dominican Republic, 16-20 November, 2021. pp. 3784–3803. Association for Computational Linguistics (2021). https://doi.org/10.18653/V1/2021.FINDINGS-EMNLP.320, https://doi.org/10.18653/v1/2021. ndings-emnlp.320|
|10.|Tang, Y., Yang, Y.: Multihop-rag: Benchmarking retrieval-augmented generation for multi-hop queries (2024)|
|11.|Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models (2023)|
|12.|Voyage AI: Voyage ai cutting-edge embedding and rerankers, https://www.voyageai.com|
|13.|Xiao, S., Liu, Z., Zhang, P., Muennighoff, N., Lian, D., Nie, J.Y.: C-pack: Packaged resources to advance general chinese embedding (2024)| | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# Multi-Meta-RAG 9
Appendix A Metadata Extraction Prompt Template
Given the question, extract the metadata to filter the database about article sources. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
report chunk embeddings might not yield optimal results. We demonstrate this multi-hop retrieval process in Figure 1.
However, existing RAG benchmarks, such as RGB (Chen et al., 2023) and RECALL (Liu et al., 2023), mainly evaluate a simple case where the answer of a query can be retrieved and solved using one single piece of evidence. None of these benchmarks assess the retrieval and reasoning capability of LLMs for complex multi-hop queries. To address this gap and make RAG benchmarking more closely resemble real-world scenarios, in this paper, we introduce MultiHop-RAG. To our knowledge, MultiHop-RAG is one of the first RAG datasets focusing specifically on multi-hop queries.
Based on the RAG queries commonly encountered in real-world scenarios, we first categorize multi-hop queries into four types: Inference query, Comparison query, Temporal query, and Null query. The first three types — Inference, Comparison, and Temporal — require the retrieval and analysis of evidence from multiple sources, encompassing tasks like inferring relationships, comparing data points, and sequencing events over time. The Null query represents a scenario where the query cannot be derived from the knowledge base. This category is crucial for assessing whether an LLM might hallucinate an answer to a multi-hop query when the retrieved text lacks relevance.
We construct our RAG knowledge base using a collection of news articles. Using GPT-4 as a data generator, we then take an extensive procedure to construct a diverse set of multi-hop queries, each requiring the retrieval and reasoning over multiple documents. An example of query construction is shown in Table 1. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# Table 9: Comparison Query Generation Prompting
Please create a time-sensitive comparison question using metadata and excerpts from multiple news articles. That is to compare the consistency or sequence of reports on similar topics at multiple different time points. If it is to compare the consistency, please clearly mention the news source and time in the question using <time frame>. If it is to compare sequences of reports, just clearly mention the news source and do not mention the timeline. Utilize the following keywords provided in the <key set> to construct the question. The correct answer should based on the factual excerpts and is only one word.
Examples
Your time-sensitive comparison question:
# Table 10: Temporal Query Generation Prompting
A multi-hop question is a query requiring multiple inferential leaps or accessing several pieces of information from different locations or sources to arrive at an answer. Considering you have read at least two news articles on <entity>, construct a multi-hop question that incorporates all the news sources. The source of the news should be stated in the question. Also, ensure that the answer to the question is a single word/entity. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
Avoid stopwords.
The sources can only be from the list: ['Yardbarker', 'The Guardian', 'Revyuh Media', 'The Independent - Sports', 'Wired', 'Sport Grill', 'Hacker News', 'Iot Business News', 'Insidesport', 'Sporting News', 'Seeking Alpha', 'The Age', 'CB-SSports.com', 'The Sydney Morning Herald', 'FOX News - Health', 'Science News For Students', 'Polygon', 'The Independent - Life and Style', 'FOX News - Entertainment', 'The Verge', 'Business Line', 'The New York Times', 'The Roar | Sports Writers Blog', 'Sportskeeda', 'BBC News - Entertainment & Arts', 'Business World', 'BBC News - Technology', 'Essentially Sports', 'Mashable', 'Advanced Science News', 'TechCrunch', 'Financial Times', 'Music Business Worldwide', 'The Independent - Travel', 'FOX News - Lifestyle', 'TalkSport', 'Yahoo News', 'Scitechdaily | Science Space And Technology News 2017', 'Globes English | Israel Business Arena', 'Wide World Of Sports', 'Rivals', 'Fortune', 'Zee Business', 'Business Today | Latest Stock Market And Economy News India', 'Sky Sports', 'Cnbc | World Business News Leader', 'Eos: Earth And Space Science News', 'Live Science: The Most Interesting Articles', 'Engadget']
Examples to follow:
Question: Who is pe individual associated wip pe cryptocurrency industry facing a criminal trial on fraud and conspiracy charges, as reported by bop The Verge and TechCrunch, and is accused by prosecutors of committing fraud for personal gain?
Answer: {'source': {'$in': ['The Verge', 'TechCrunch']}}
Question: After pe TechCrunch report on October 7, 2023, concerning Dave Clark’s comments on Flexport, and pe subsequent TechCrunch article on October 30, 2023, regarding Ryan Petersen’s actions at Flexport, was pere a change in pe nature of pe events reported?
Answer: {'source': {'$in': ['TechCrunch']}, 'published_at': '$in': ['October 7, 2023', 'October 30, 2023']}
Question: Which company, known for its dominance in pe e-reader space and for offering exclusive invite-only deals during sales events, faced a stock decline due to an antitrust lawsuit reported by 'The Sydney Morning Herald' and discussed by sellers in a 'Cnbc | World Business News Leader' article?
Answer: {'source': {'$in': ['The Sydney Morning Herald', 'Cnbc | World Business News Leader']}
If you detect multiple queries, return the answer for the first. Now it is your turn:
Question: <query>
Answer: | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# arXiv:2405.07437v1 [cs.CL] 13 May 2024
Evaluation of Retrieval-Augmented Generation: A Survey
Hao Yu1,2, Aoran Gan3, Kai Zhang3, Shiwei Tong1†, Qi Liu3, and Zhaofeng Liu1
1 Tencent Company
2 McGill University
3 State Key Laboratory of Cognitive Intelligence, University of Science and Technology of China
hao.yu2@mail.mcgill.ca
{shiweitong†,zhaofengliu}@tencent.comgar@mail.ustc.edu.cn
{kkzhang08,qiliuql}@ustc.edu.cn
Abstract. Retrieval-Augmented Generation (RAG) has emerged as a pivotal innovation in natural language processing, enhancing generative models by incorporating external information retrieval. Evaluating RAG systems, however, poses distinct challenges due to their hybrid structure and reliance on dynamic knowledge sources. We consequently enhanced an extensive survey and proposed an analysis framework for benchmarks of RAG systems, RGAR (Retrieval, Generation, Additional Requirement), designed to systematically analyze RAG benchmarks by focusing on measurable outputs and established truths. Specifically, we scrutinize and contrast multiple quantifiable metrics of the Retrieval and Generation component, such as relevance, accuracy, and faithfulness, of the internal links within the current RAG evaluation methods, covering the possible output and ground truth pairs. We also analyze the integration of additional requirements of different works, discuss the limitations of current benchmarks, and propose potential directions for further research to address these shortcomings and advance the field of RAG evaluation. In conclusion, this paper collates the challenges associated with RAG evaluation. It presents a thorough analysis and examination of existing methodologies for RAG benchmark design based on the proposed RGAR framework.
# 1 Introduction
Retrieval-Augmented Generation (RAG) [29] represents a significant advancement in natural language processing by enhancing the performance of generative models through integrating information retrieval techniques. It addresses a critical challenge faced by standalone generative models: the tendency to produce responses that, while plausible, may not be grounded in facts. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
By retrieving relevant information from external sources,† Corresponding Author
Paper Homepage: https://github.com/YHPeter/Awesome-RAG-Evaluation | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
2 Hao Yu, Aoran Gan, Kai Zhang, Shiwei Tong†, Qi Liu, and Zhaofeng Liu
RAG significantly reduces the incidence of hallucinations [20] or factually incorrect generative outputs, thereby improving the content’s reliability and richness. [56] This fusion of retrieval and generation capabilities enables the creation of responses that are not only contextually appropriate but also informed by the most current and accurate information available, making RAG a development in the pursuit of more intelligent and versatile language models [56,52].
|Retrieval|Generation|Ground Truth|
|---|---|---|
|Wob Search Enging|Query|Prompt|
|Query|BM25|Doctu ScorsRelevami DelevaniDocs 8 CompictcPrompt|
|KNNANN|Ococc|System PromptPrompt Skills|
| | |Docs Candidates|
|Wikinedia|RLSDONSC|GamoieResdanse|
|HF Datase|Post|Processing|
| | |Qutput|
|Indexing|IV . Inferencing|Labei|
Fig. 1: The structure of the RAG system with retrieval and generation components and corresponding four phrases: indexing, search, prompting and inferencing. The pairs of EOs and GTs are highlighted in red and green, with brown dashed arrows.
Numerous studies of RAG systems have emerged from various perspectives since the advent of pre-trained language models [14]. The RAG system comprises two primary components: Retrieval and Generation. The retrieval component aims to extract relevant information from a vast array of external knowledge sources. It involves two main phases, indexing and searching. Indexing organizes documents to facilitate efficient retrieval, using either inverted indexes for sparse retrieval or dense vector encoding for dense retrieval [14,11,24]. The searching utilizes these indexes to fetch relevant documents on the user’s query, often incorporating the optional rerankers [4,34,5,43] to refine the ranking of the retrieved documents. The generation component utilizes the retrieved content to formulate coherent and contextually relevant responses with the prompting and inferencing phases. The input for language models is formulated through prompting, integrating the query from the retrieval phase. Methods like Chain of Thought (CoT) [48] or Rephrase and Respond (RaR) [7] guide better generation results. In the inferencing step, Large Language Models (LLMs) interpret the prompted input to generate accurate and in-depth responses that align with the query’s intent and integrate the extracted information [30,8]. The RAG structure is explained in detail in Appendix A. Figure 1 illustrates the structure of the RAG systems as mentioned.
| Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
The significance of evaluating RAG is escalating in tandem with the advancement of RAG-specific methodologies. On the one hand, RAG is a complex system intricately tied to specific requirements and language models, resulting in a variety of evaluation | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# Evaluation of Retrieval-Augmented Generation: A Survey
Methods, indicators, and tools, particularly given the black-box LLM generation. Evaluating RAG systems thus involves considering quite a few specific components and the complexity of overall system assessment. On the other hand, the complexity of RAG systems is further compounded by the dynamic external database and the various downstream tasks, such as content creation or open domain question answering [14,53]. These challenges necessitate the development of comprehensive evaluation metrics that can effectively capture the interplay between retrieval accuracy and generative quality [2,6]. To clarify the elements further, we conducted this survey on RAG evaluation to address the current gaps in the area, which differs from the prior RAG surveys [57,14,21] that predominantly collected specific RAG methods or data. We have compiled 12 distinct evaluation frameworks, encompassing a range of aspects of the RAG system. We conduct a comparative analysis and synthesize the specific evaluation methods of various components, focusing on aspects such as accuracy, faithfulness, and relevance. We also discuss the constraints of the existing methodology and the prospects for future RAG evaluations. We hope to provide the readers with a comprehensive understanding of the RAG evaluation.
For this paper, we contribute in the following aspects:
1. Challenge of Evaluation: This is the first work that summarizes and classifies the challenges in evaluating RAG systems through the structure of RAG systems, including three parts retrieval, generation, and the whole system.
2. Analysis Framework: Based on the challenges, we propose an analysis framework (RGAG) for RAG benchmarks, which is designed to navigate the unique complexities inherent to RAG systems, offering a fundamental methodology for assessing their efficacy across many facets.
3. RAG Benchmark Analysis: With the help of the RGAG framework, we provide a comprehensive analysis of existing RAG benchmarks, highlighting their strengths and limitations and proposing recommendations for future developments in RAG system evaluation.
# Challenges in Evaluating RAG Systems
Evaluating hybrid RAG systems entails evaluating retrieval, generation and the RAG system as a whole. These evaluations are multifaceted, requiring careful consideration and analysis. Each of them encompasses specific difficulties that complicate the development of a comprehensive evaluation framework and benchmarks for RAG systems.
Retrieval: The retrieval component of RAG systems is critical for fetching relevant information from external knowledge sources that inform the generation process. One primary challenge in evaluating this component is the dynamic and vast nature of potential knowledge bases, ranging from structured databases to the entire web. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
This vastness requires evaluation metrics that can effectively measure the precision, recall, and relevance of retrieved documents in the context of a given query [43,27]. Moreover, the temporal aspect of information, where the relevance and accuracy of data can change over time, adds another layer of complexity to the evaluation process [5]. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
Additionally, | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
4 Hao Yu, Aoran Gan, Kai Zhang, Shiwei Tong†, Qi Liu, and Zhaofeng Liu
The diversity of information sources and the potential for retrieving misleading or low-quality information poses significant challenges in assessing the retrieval component’s effectiveness in filtering and selecting the most pertinent information [34]. On the other hand, the evaluation method of the retrieval component is also a challenge, as the traditional evaluation metrics enforce the retrieval system focus on higher TopK recall, instead of the useful information carried through one query.
Generation The generation component, currently powered by LLMs, produces coherent and contextually appropriate responses based on the retrieved content. A significant challenge here lies in evaluating the faithfulness and accuracy of the generated content to the input data. This involves not only assessing the factual correctness of responses but also their relevance to the original query and the coherence of the generated text [58,41]. The subjective nature of certain tasks, such as creative content generation or open-ended question answering, further complicates the evaluation, as it introduces variability in what constitutes a ‘correct’ or ‘high-quality’ response [40].
RAG System as a Whole Evaluating the whole RAG system introduces additional complexities. The interplay between the retrieval and generation components means that the entire system’s performance cannot be fully understood by evaluating each component in isolation [41,13]. The system needs to be assessed on its ability to leverage retrieved information effectively to improve response quality, which involves measuring the added value of the retrieval component to the generative process. Furthermore, practical considerations such as response latency, robustness against misinformation, and the ability to handle ambiguous or complex queries are also crucial for evaluating the system’s overall effectiveness and usability [34,5].
# 3 RGAR: Analysis Framework for Evaluation
Evaluating the target shift from traditional absolute numeric metrics to multi-source and multi-target generation evaluation, along with the intricate interplay between retrieval and generation components, poses significant challenges. Searches in a dynamic database may lead to misleading results or contradicting the facts. Diverse and comprehensive datasets that accurately reflect real-world scenarios are crucial. Challenges also arise in the realm of metrics, encompassing generative evaluation criteria for distinct downstream tasks, human preferences, and practical considerations within the RAG system. Most prior benchmarks predominantly tackle one or several aspects of the RAG assessment but lack a comprehensive, holistic analysis. To provide a better understanding of RAG benchmarks, we propose an analysis framework named RGAR (Retrieval, Generation, and Additional Requirement). It takes into account the Target, Dataset, and Metric respectively. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
The Target module is intended to determine the evaluation direction. The Dataset module facilitates the comparison of various data constructions in RAG benchmarks. The final module, Metrics, introduces the metrics that correspond to specific targets and datasets used during evaluation. Overall, it is designed to provide a systematic methodology for assessing the effectiveness of RAG systems across various aspects by covering all possible pairs between the “Evaluable Outputs” (EOs) and “Ground Truths” (GTs). In the following | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
Do not answer this question directly. Just give me the question:
# Table 11: Null Query Generation Prompting | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# Evaluation of Retrieval-Augmented Generation: A Survey
In this section, we will explain thoroughly the framework and utilize it for introducing and comparing the RAG benchmarks.
# RGAR Framework
|Output|Target|Ground Truth|
|---|---|---|
|Query|Retrieval Relevant Docs|Relevance Query|
|Relevant Docs| |Docs Candidates|
|Response|Generation Query|Relevance|
|Response| |Faithfulness|
|Response| |Correctness|
|Response| |Sample Response|
|Response| |Additional Requirements|
|Output| |Label|
Fig. 2: The Target modular of RGAR framework.
| Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# Evaluation Target (What to Evaluate?)
The combination of EOs and GTs in the RAG system can generate all possible targets, which is the fundamental concept of the RGAR framework (as shown in Figure 1). Once identified, these targets can be defined based on a specific pair of EOs or EO with GT, as illustrated in Figure 2, and used to analyze all aspects of current RAG benchmarks.
Retrieval For the evaluation of the retrieval component, the EOs are the relevant documents depending on the query. Then we can construct two pairwise relationships for the retrieval component, which are Relevant Documents ↔ Query, Relevant Documents ↔ Documents Candidates.
- Relevance (Relevant Documents ↔ Query) evaluates how well the retrieved documents match the information needed expressed in the query. It measures the precision and specificity of the retrieval process.
- Accuracy (Relevant Documents ↔ Documents Candidates) assesses how accurate the retrieved documents are in comparison to a set of candidate documents. It is a measure of the system’s ability to identify and score relevant documents higher than less relevant or irrelevant ones. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# Hao Yu, Aoran Gan, Kai Zhang, Shiwei Tong†, Qi Liu, and Zhaofeng Liu
Generation The similar pairwise relations for the generation components are listed below. The EOs are the generated text and phrased structured content. Then we need to compare these EOs with the provided GTs and labels.
- Relevance (Response ↔ Query) measures how well the generated response aligns with the intent and content of the initial query. It ensures that the response is related to the query topic and meets the query’s specific requirements.
|Category|Framework|Time|Raw Targets|Retrieval|Generation|
|---|---|---|---|---|---|
|Tool|TruEra RAG Triad [45]|2023.10|Answer Relevance|LLM as Judge|LLM as Judge|
|Tool|LangChain Bench. [27]|2023.11|Faithfulness|Accuracy|LLM as Judge|
|Tool|Databricks Eval [28]|2023.12|Readability|-|LLM as Judge|
|Benchmark|RAGAs [13]|2023.09|Answer Relevance|LLM as Judge|LLM as Judge|
|Benchmark|RECALL [33]|2023.11|Response Quality|-|BLEU, ROUGE-L|
|Benchmark|ARES [41]|2023.11|Answer Faithfulness|LLM + Classifier|LLM + Classifier|
|Benchmark|RGB [5]|2023.12|Noise Robustness|-|Accuracy|
|Benchmark|MultiHop-RAG [43]|2024.01|Retrieval Quality|MAP, MRR, Hit@K|LLM as Judge|
|Benchmark|CRUD-RAG [34]|2024.02|CREATE, READ, UPDATE, DELETE|-|ROUGE, BLEU, RAGQuestEval|
|Benchmark|MedRAG [49]|2024.02|Accuracy|-|Accuracy|
|Benchmark|FeB4RAG [47]|2024.02|Correctness|-|Human Evaluation|
|Benchmark|CDQA [50]|2024.03|Accuracy|-|F1 of Tokens|
|Research|FiD-Light [18]|2023.07|Latency|-|-|
|Research|Diversity Reranker [4]|2023.08|Diversity|-|-|
Table 1: The evaluating targets and corresponding metrics across various frameworks for evaluating RAG systems. The presentation distinguishes between the core areas of Retrieval and Generation considered in the evaluation. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
The different aspects of the evaluation are set as different colours in the table: Relevance, Accuracy of Retrieval and Faithfulness, Correctness and Relevance of Generation. The consideration of the Additional Requirements beyond the retrieval and generation component is also collected. Noted that quite a few of the works employed multiple methods or evaluated multiple aspects simultaneously. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# Evaluation of Retrieval-Augmented Generation: A Survey
- Faithfulness (Response ↔ Relevant Documents) evaluates if the generated response accurately reflects the information contained within the relevant documents and measures the consistency between generated content and the source documents.
- Correctness (Response ↔ Sample Response) Similar to the accuracy in the retrieval component, this measures the accuracy of the generated response against a sample response, which serves as a ground truth. It checks if the response is correct in terms of factual information and appropriate in the context of the query.
| Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
The targets of Retrieval and Generation components are introduced. Table 1 lists the relative work about improving and evaluating RAG and RAG benchmarks cut off in March 2024. Table 1 portrays this information, where each evaluation criterion is represented by a different colour. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
For example, FeB4RAG [47], the fourth from the last, has posited four standards based on [15] that comprise Consistency, Correctness, Clarity, and Coverage. Correctness is equivalent to accuracy in retrieval, and Consistency is tantamount to faithfulness in the generation component. While accuracy in retrieval gauges the correctness of the retrieved information, we posit that Coverage pertains to the coverage rate and is more associated with diversity. Therefore, we consider Coverage to be linked with diversity and an additional requirement in our proposed evaluation framework, which will be introduced subsequently. The remaining standard, Clarity, is also classified as an additional requirement in our proposed framework. The other tools and benchmarks are processed similarly.
Tools and benchmarks offer varying degrees of flexibility in evaluating datasets for RAG systems. Tools, which specify only evaluation targets, provide a versatile framework capable of constructing complete RAG applications and evaluation pipelines, as seen in works like [45,27,28]. Benchmarks, on the other hand, focus on different aspects of RAG evaluation with specific emphasis on either retrieval outputs or generation targets. For instance, RAGAs and ARES assess the relevance of retrieval documents, while RGB and MultiHop-RAG [5,43] prioritize accuracy, necessitating comparison with GTs. All benchmarks consider generation targets due to their critical role in RAG systems, though their focus areas vary.
# Additional Requirement
In addition to evaluating the two primary components outlined, a portion of the works also addressed some additional requirements of RAG (Black and Ltatic targets in the Table 2). The requirements are as follows:
- Latency [18,27] measures how quickly the system can find information and respond, crucial for user experience.
- Diversity [4,27] checks if the system retrieves a variety of relevant documents and generates diverse responses.
- Noise Robustness [5] assesses how well the system handles irrelevant information without affecting response quality.
- Negative Rejection [5] gauges the system’s ability to refrain from providing a response when the available information is insufficient.
- Counterfactual Robustness [5] evaluates the system’s capacity to identify and disregard incorrect information, even when alerted about potential misinformation.
- More: For more human preferences considerations, there can be more additional requirements, such as readability [47,28], toxicity, perplexity [28], etc. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# 8 Hao Yu, Aoran Gan, Kai Zhang, Shiwei Tong†, Qi Liu, and Zhaofeng Liu
For the exception, CRUD-RAG [34] introduces a comprehensive benchmark addressing the broader spectrum of RAG applications beyond question-answering, categorized into Create, Read, Update, and Delete scenarios. This benchmark evaluates RAG systems across diverse tasks, including text continuation, question answering, hallucination modification, and multi-document summarization, each corresponding to CRUD actions. It emphasizes the evaluation of all RAG components, including retrieval models and external knowledge base construction, offering insights for optimizing RAG technology across different scenarios.
# 3.2 Evaluation Dataset (How to evaluate?)
|Benchmark|Dataset|
|---|---|
|RAGAs [13]|WikiEval|
|RECALL [33]|EventKG [17], UJ [19]|
|ARES [41]|NQ [25], Hotpot [51], FEVER [44], WoW [10], MultiRC [9], ReCoRD [54]|
|RGB [5]|Generated (Source: News)|
|MultiHop-RAG [43]|Generated (Source: News)|
|CRUD-RAG [34]|Generated (Source: News)UHGEval [31]|
|MedRAG [49]|MIRAGE|
|FeB4RAG [47]|FeB4RAG, BEIR [23]|
|CDQA [50]|Generations (Source: News), Labeller|
Table 2: The evaluation datasets used for each benchmark. The dataset without citation was constructed by the benchmark itself.
In Table 2, distinct benchmarks employ varying strategies for dataset construction, ranging from leveraging existing resources to generating entirely new data tailored for specific evaluation aspects. Several benchmarks draw upon the part of KILT (Knowledge Intensive Language Tasks) benchmark [37] (Natural Questions [25], HotpotQA [51], and FEVER [44]) and other established datasets such as SuperGLUE [46] (MultiRC [9], and ReCoRD [54]) [41]. However, the drawback of using such datasets can’t solve the challenges in dynamic real-world scenarios. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
A similar situation can be observed in WikiEval constructed by RAGAs [13], which was generated from Wikipedia pages post 2022.
The advent of powerful LLMs has revolutionized the process of dataset construction. With the ability to design queries and ground truths for specific evaluation targets using these frameworks, authors can now create datasets in the desired format with ease. Benchmarks like RGB, MultiHop-RAG, CRUD-RAG, and CDQA [5,43,34,50] have taken this approach further by building their own datasets using online news articles to test RAG systems’ ability to handle real-world information beyond the training data of LM frameworks. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# 3.3 Evaluation Metric (How to quantify?)
Navigating the intricate terrain of evaluating RAG systems necessitates a nuanced understanding of the metrics that can precisely quantify the evaluation targets. However, creating evaluative criteria that align with human preferences and address practical considerations is challenging. Each component within the RAG systems requires a tailored evaluative approach that reflects its distinct functionalities and objectives.
Retrieval Metrics
Various targets can be evaluated with various metrics that correspond to the given datasets. This section will introduce several commonly used metrics for retrieval and generation targets. The metrics for additional requirements can also be found in these commonly used metrics. The more specifically designed metrics can be explored in the original paper via Table 1 as a reference.
For the retrieval evaluation, the focus is on metrics that can accurately capture the relevance, accuracy, diversity, and robustness of the information retrieved in response to queries. These metrics must not only reflect the system’s precision in fetching pertinent information but also its resilience in navigating the dynamic, vast, and sometimes misleading landscape of available data. The deployment of metrics like Misleading Rate, Mistake Reappearance Rate, and Error Detection Rate within the benchmark underscores a heightened awareness of RAG systems’ inherent intricacies. The integration of MAP@K, MRR@K, and Tokenization with F1 into benchmarks like mirrors a deepening comprehension of traditional retrieval’s multifaceted evaluation. While the also emphasizes that this ranking-based evaluation methodology is not unsuitable for the RAG system, and should have more RAG-specific retrieval evaluation metrics. These metrics not only capture the precision and recall of retrieval systems but also account for the diversity and relevance of retrieved documents, aligning with the complex and dynamic nature of information needs in RAG systems. The introduction of LLMs as evaluative judges, as seen in , further underscores the adaptability and versatility of retrieval evaluation, offering a comprehensive and context-aware approach to assessing retrieval quality.
Non-Rank Based Metrics
Non-rank-based metrics often assess binary outcomes—whether an item is relevant or not—without considering the position of the item in a ranked list. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# Query:
Which platform is at the center of discussions in articles from Music Business Worldwide, Polygon, and FOX News - Health, concerning the policing of AI-driven voice replication, the debate over "reaction" content, and being the most used app overnight by young people?
# Answer:
YouTube
# Evidence List:
|Title|Sony Music’s artists aren’t involved in YouTube’s new voice-cloning AI experiment.|
|---|---|
|Source|Music Business Worldwide|
|Published Time|2023-11-23T18:48:48+00:00|
|Fact|During this period of discussion, YouTube has made a number of positive announcements regarding the biggest issue for any rightsholder regarding AI-driven voice replication of artists: their ability to police it.|
|Title|YouTube demonetizes popular content creator SSSniperwolf after doxxing accusations|
|---|---|
|Source|Polygon|
|Published Time|2023-10-25T18:18:06+00:00|
|Fact|The debate over "reaction" content on YouTube has been brewing for years, but a recent incident between two creators has refueled the urgency of the conversation.|
|Title|Cell phone shocker as 97% of kids use their device during school hours and beyond, says study|
|---|---|
|Source|FOX News - Health|
|Published Time|2023-10-01T09:05:26+00:00|
|Fact|Overnight phone use was primarily spent engaging with the same media, although YouTube appeared to be the longest-running app because videos were often left playing during the night.|
# Query:
Did the Cnbc | World Business News Leader report on Nike’s net income and the article from The Age on the 10-year Treasury yield both report a decrease in their respective financial metrics?
| Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
Notice, that the following formula is just one format of these metrics, the definition of each metric may vary by the different evaluating tasks.
Accuracy
Accuracy = T P + T N / T P + T N + F P + F N
where T P is pe number of true positives, T N is pe number of true negatives, F P is pe number of false positives, and F N is pe number of false negatives. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
10 Hao Yu, Aoran Gan, Kai Zhang, Shiwei Tong†, Qi Liu, and Zhaofeng Liu
- Precision is the fraction of relevant instances among the retrieved instances.
|Precision =|TP|
|---|---|
| |TP + FP|
where TP represents true positives and FP represents false positives.
- Recall at k (Recall@k) is the fraction of relevant instances that have been retrieved over the total amount of relevant instances, considering only the top k results.
|Recall@k =||RD ∩ Topkd|| |
|---|---|---|
| ||RD|| |
where RD is the relevant documents, and Topkd is the top-k retrieved documents.
Rank-Based Metrics Rank-based metrics evaluate the order in which relevant items are presented, with higher importance placed on the positioning of relevant items at the top of the ranking list.
- Mean Reciprocal Rank (MRR) is the average of the reciprocal ranks of the first correct answer for a set of queries.
|MRR =|Σ|Q||1|
|---|---|---|
| |Σi=1 ranki| |
where |Q| is the number of queries and ranki is the rank position of the first relevant document for the i-th query.
- Mean Average Precision (MAP) is the mean of the average precision scores for each query.
|MAP =|Σ|Q||ΣP(k) × rel(k)|
|---|---|---|
| |Σq=1||Q||
| | |Σq=1 |relevant documentsq ||
where P(k) is the precision at cutoff k in the list, rel(k) is an indicator function equaling 1 if the item at rank k is a relevant document, 0 otherwise, and n is the number of retrieved documents.
Generation Metrics In the realm of generation, evaluation transcends the mere accuracy of generated responses, venturing into the quality of text in terms of coherence, relevance, fluency, and alignment with human judgment. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
This necessitates metrics that can assess the nuanced aspects of language production, including factual correctness, readability, and user satisfaction with the generated content. The traditional metrics like BLEU, ROUGE, and F1 Score continue to play a crucial role, emphasizing the significance of precision and recall in determining response quality. Yet, the advent of metrics such as Misleading Rate, Mistake Reappearance Rate, and Error Detection Rate highlights an evolving understanding of RAG systems’ distinct challenges [33]. The evaluation done by humans is still a very significant standard to compare the performance of generation models with one another or with the ground truth. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# Evaluation of Retrieval-Augmented Generation: A Survey
The approach of employing LLMs as evaluative judges [58] is a versatile and automatic method for quality assessment, catering to instances where traditional ground truths may be elusive [13]. This methodology benefits from employing prediction-powered inference (PPI) and context relevance scoring, offering a nuanced lens through which LLM output can be assessed. [41] The strategic use of detailed prompt templates ensures a guided assessment aligned with human preferences, effectively standardizing evaluations across various content dimensions [1]. This shift towards leveraging LLMs as arbiters marks a significant progression towards automated and context-responsive evaluation frameworks, enriching the evaluation landscape with minimal reliance on reference comparisons.
# ROUGE
Recall-Oriented Understudy for Gisting Evaluation (ROUGE) [32] is a set of metrics designed to evaluate the quality of summaries by comparing them to human-generated reference summaries. The variants of ROUGEs measure the overlap of n-grams (ROUGE-N, ROUGGE-W), word subsequences (ROUGE-L, ROUGGE-S), and word pairs between the system-generated summary and the reference summaries. ROUGE can be indicative of the content overlap between the generated text and the reference text.
# BLEU
Bilingual Evaluation Understudy (BLEU) [36] is a metric for evaluating the quality of machine-translated text against one or more reference translations. BLEU calculates the precision of n-grams in the generated text compared to the reference text and then applies a brevity penalty to discourage overly short translations. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
BLEU has been applied to other generation tasks. Despite its popularity, BLEU has limitations, such as not accounting for the fluency or grammaticality of the generated text.
# BertScore
BertScore [55] leverages the contextual embedding from pre-trained transformers like BERT to evaluate the semantic similarity between generated text and reference text. BertScore computes token-level similarity using contextual embedding and produces precision, recall, and F1 scores. Unlike n-gram-based metrics, BertScore captures the meaning of words in context, making it more robust to paraphrasing and more sensitive to semantic equivalence.
# LLM as a Judge
Using LLMs as judges for evaluating generated text is a more recent approach. [58] In this method, LLMs are used to score the generated text based on criteria such as coherence, relevance, and fluency. The LLM can be fine-tuned on human judgments to predict the quality of unseen text or used to generate evaluations in a zero-shot or few-shot setting. This approach leverages the LLM’s understanding of language and context to provide a more nuanced assessment of text quality. For instance, [1] illustrates how providing LLM judges with detailed scoring guidelines, such as a scale from 1 to 5, can standardize the evaluation process. This methodology encompasses critical aspects of content assessment, including coherence, relevance, fluency, coverage, diversity, and detail - both in the context of answer evaluation and query formulation. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
12 Hao Yu, Aoran Gan, Kai Zhang, Shiwei Tong†, Qi Liu, and Zhaofeng Liu
Additional Requirements These additional requirements, such as latency, diversity,
noise robustness, negative rejection, and counterfactual robustness, are used to ensure
the practical applicability of RAG systems in real-world scenarios aligned with human
preference. This section delves into the metrics used for evaluating these additional
requirements, highlighting their significance in the comprehensive assessment of RAG
systems.
Latency measures the time taken by the RAG system to retrieve relevant information
and generate a response. It is a critical factor for user experience, especially in interactive applications such as chatbots or search engines [18].
Single Query Latency:
The average time taken to process a single query, including both retrieval and generating phases.
Diversity evaluates the variety and breadth of information retrieved and generated by
the RAG system. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
It ensures that the system can provide a wide range of perspectives
and avoid redundancy in responses [4].
Cosine Similarity / Cosine Distance:
The cosine similarity/distance calculates embeddings of retrieved documents or generated responses. Lower cosine similarity scores indicate higher diversity, suggesting that the system can retrieve or generate a broader spectrum of information.
Noise Robustness measures the RAG system’s ability to handle irrelevant or misleading information without compromising the quality of the response [33]. The metrics Misleading Rate and Mistake Reappearance Rate are described in [33], providing detailed descriptions tailored to the specific dataset and experimental setup.
Negative Rejection evaluates the system’s capability to withhold responses when the available information is insufficient or too ambiguous to provide an accurate answer [5].
Rejection Rate:
The rate at which the system refrains from generating a response.
Counterfactual Robustness assesses the system’s ability to identify and disregard incorrect or counterfactual information within the retrieved documents [34].
Error Detection Rate:
The ratio of counterfactual statements are detected in retrieved information.
4 Discussion
Evaluating Retrieval-Augmented Generation (RAG) systems comprises multifaceted
challenges, arising from their dual reliance on retrieving accurate, relevant information
and generating coherent responses that meet user expectations. This survey has highlighted several key points of consideration, capturing the breadth and depth of evaluations necessary for advancing RAG technologies.
Within the evaluation target aspect, it’s evident that traditional question-answering
(QA) setups remain effective for evaluating generative components, particularly in how | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# Evaluation of Retrieval-Augmented Generation: A Survey
Well they align with human preferences for clarity, relevance, and factual accuracy. However, recent strategies, such as CRUD-based assessments, offer novel angles by scrutinizing RAG systems’ interactive capabilities with dynamic information environments [34]. These methodologies underscore the necessity for RAG evaluations to evolve beyond static benchmarks, mirroring real-world scenarios where information is continuously updated and queries are not strictly fact-based but exploratory or conversational.
On the dataset front, the challenge of devising a "one-size-fits-all" dataset is pronounced, given the highly task-specific nature of RAG systems. Unique datasets, meticulously crafted to test specific facets of RAG performance, are indispensable. While this approach ensures thorough, targeted evaluation, it also magnifies the effort and resources required for comprehensive testing. The divergence of datasets, ranging from news articles to structured databases, reflects the adaptability required of RAG systems but also signifies a formidable barrier to streamlined evaluation [49,50].
When it comes to metrics, the use of LLMs as automatic evaluative judges signifies a burgeoning trend, promising versatility and depth in generative outputs with reasoning on a large scale compared to human evaluation. However, using LLMs as judges for chatbot responses presents challenges in aligning with human judgment, establishing effective grading scales, and applying consistent evaluation across varied use cases. The determination of correctness, clarity, and richness can differ between automated and human assessments. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
Moreover, the effectiveness of example-based scoring can vary, and there’s no universally applicable grading scale, complicating the standardization of LLM as a judge. [28]
Future directions in RAG evaluation should focus on developing more adaptive, context-aware benchmarks that accurately reflect the dynamic, information-rich environments these systems are designed to navigate. Such efforts could include simulating real-time information updates in evaluation datasets or incorporating user feedback loops into assessment methodologies. Additionally, exploring more nuanced metrics that can capture the subtlety of human language comprehension and generation—beyond sheer accuracy or relevance—will be crucial. Efforts to codify these advancements into standardized evaluation frameworks would significantly bolster the field, providing clearer benchmarks for progress and more directly aligning RAG system advancements with user needs and societal impacts.
# Conclusion
This survey has systematically explored the complex landscape of evaluating Retrieval-Augmented Generation (RAG) systems, highlighting the multifaceted challenges inherent in assessing their performance. Through the proposed RGAR analysis framework, we have delineated a structured approach to dissecting the intricacies of RAG evaluations, focusing on the retrieval, generation, and additional requirements that underpin these systems. Our comprehensive analysis underscores the necessity for targeted benchmarks that reflect the dynamic interplay between retrieval accuracy and generative quality, as well as the practical considerations crucial for real-world applications. By identifying gaps in current methodologies and suggesting future research directions. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
14 Hao Yu, Aoran Gan, Kai Zhang, Shiwei Tong†, Qi Liu, and Zhaofeng Liu
this survey aims to pave the way for more nuanced, effective, and user-aligned evaluations of RAG systems, ultimately contributing to the advancement of natural language processing technologies that are both intelligent and versatile. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# Answer:
Yes
# Evidence List:
|Title|Nike misses revenue expectations for the first time in two years, beats on earnings and gross margin|
|---|---|
|Source|Cnbc | World Business News Leader|
|Published Time|2023-09-28T20:31:00+00:00|
|Fact|The company’s reported net income for the three-month period that ended August 31 was $1.45 billion, or 94 cents per share, compared with $1.47 billion, or 93 cents per share, a year earlier.|
|Title|ASX set to open higher as Wall Street rebounds; $A rises|
|---|---|
|Source|The Age|
|Published Time|2023-10-04T21:01:01+00:00|
|Fact|The yield on the 10-year Treasury, which is the centrepiece of the bond market, pulled back from its highest level since 2007, down to 4.73 per cent from 4.80 per cent late on Tuesday.| | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# References
|1.|Balaguer, A., Benara, V., Cunha, R.L.d.F., Filho, R.d.M.E., Hendry, T., Holstein, D., Marsman, J., Mecklenburg, N., Malvar, S., Nunes, L.O., Padilha, R., Sharp, M., Silva, B., Sharma, S., Aski, V., Chandra, R.: RAG vs Fine-tuning: Pipelines, Tradeoffs, and a Case Study on Agriculture. Tech. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
rep. (Jan 2024), http://arxiv.org/abs/2401.08406, arXiv:2401.08406 [cs] type: article|
|---|---|
|2.|Barnett, S., Kurniawan, S., Thudumu, S., Brannelly, Z., Abdelrazek, M.: Seven failure points when engineering a retrieval augmented generation system (Jan 2024). https://doi.org/10.48550/ARXIV.2401.05856|
|3.|Besta, M., Blach, N., Kubicek, A., Gerstenberger, R., Podstawski, M., Gianinazzi, L., Gajda, J., Lehmann, T., Niewiadomski, H., Nyczyk, P., Hoefler, T.: Graph of thoughts: Solving elaborate problems with large language models. Proceedings of the AAAI Conference on Artificial Intelligence 2024 (AAAI’24) (Aug 2023). https://doi.org/10.48550/ARXIV.2308.09687|
|4.|Blagojevic, V.: Enhancing RAG Pipelines in Haystack: Introducing DiversityRanker and LostInTheMiddleRanker (Aug 2023), https://towardsdatascience.com/enhancing-rag-pipelines-in-haystack-45f14e2bc9f5|
|5.|Chen, J., Lin, H., Han, X., Sun, L.: Benchmarking large language models in retrieval-augmented generation (Sep 2023). https://doi.org/10.48550/ARXIV.2309.01431|
|6.|Cuconasu, F., Trappolini, G., Siciliano, F., Filice, S., Campagnano, C., Maarek, Y., Tonellotto, N., Silvestri, F.: The power of noise: Redefining retrieval for rag systems (Jan 2024). https://doi.org/10.48550/ARXIV.2401.14887|
|7.|Deng, Y., Zhang, W., Chen, Z., Gu, Q.: Rephrase and respond: Let large language models ask better questions for themselves (Nov 2023). https://doi.org/10.48550/ARXIV.2311.04205|
|8.|Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In: Burstein, J., Doran, C., Solorio, T. (eds.) Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers). pp. 4171–4186. Association for Computational Linguistics, Minneapolis, Minnesota (Jun 2019). https://doi.org/10.18653/v1/N19-1423, https://aclanthology.org/N19-1423|
|9.|DeYoung, J., Jain, S., Rajani, N.F., Lehman, E., Xiong, C., Socher, R., Wallace, B.C.: Eraser: A benchmark to evaluate rationalized nlp models|
|10.|Dinan, E., Roller, S., Shuster, K., Fan, A., Auli, M., Weston, J.: Wizard of Wikipedia: Knowledge-powered conversational agents. In: Proceedings of the International Conference on Learning Representations (ICLR) (2019)|
|11.|Douze, M., Guzhva, A., Deng, C., Johnson, J., Szilvasy, G., Mazaré, P.E., Lomeli, M., Hosseini, L., Jégou, H.: The faiss library (2024)|
|12.|DuckDuckGo: DuckDuckGo — Privacy, simplified. (2024), https://duckduckgo.com//home|
|13.|Es, S., James, J., Espinosa-Anke, L., Schockaert, S.: Ragas: Automated evaluation of retrieval augmented generation (Sep 2023). https://doi.org/10.48550/ARXIV.2309.15217|
|14.|Gao, Y., Xiong, Y., Gao, X., Jia, K., Pan, J., Bi, Y., Dai, Y., Sun, J., Guo, Q., Wang, M., Wang, H.: Retrieval-Augmented Generation for Large Language Models: A Survey. Tech. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
rep. (Jan 2024), http://arxiv.org/abs/2312.10997, arXiv:2312.10997 [cs] type: article| | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# Hao Yu, Aoran Gan, Kai Zhang, Shiwei Tong†, Qi Liu, and Zhaofeng Liu
15. Gienapp, L., Scells, H., Deckers, N., Bevendorff, J., Wang, S., Kiesel, J., Syed, S., Fröbe, M., Zuccon, G., Stein, B., Hagen, M., Potpast, M.: Evaluating Generative Ad Hoc Information Retrieval. Tech. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
rep. (Nov 2023), http://arxiv.org/abs/2311.04694, arXiv:2311.04694 [cs] type: article
16. Google: Programmable Search Engine | Google for Developers (2024), https://developers.google.com/custom-search
17. Gottschalk, S., Demidova, E.: Eventkg: A multilingual event-centric temporal knowledge graph (Apr 2018). https://doi.org/10.48550/ARXIV.1804.04526
18. Hofstätter, S., Chen, J., Raman, K., Zamani, H.: FiD-Light: Efficient and Effective Retrieval-Augmented Text Generation. In: Proceedings of pe 46p International ACM SIGIR Conference on Research and Development in Information Retrieval. pp. 1437–1447. SIGIR ’23, Association for Computing Machinery, New York, NY, USA (Jul 2023). https://doi.org/10.1145/3539618.3591687, https://doi.org/10.1145/3539618.3591687
19. Huang, J., Shao, H., Chang, K.C.C., Xiong, J., Hwu, W.m.: Understanding jargon: Combining extraction and generation for definition modeling. In: Proceedings of EMNLP (2022)
20. Huang, L., Yu, W., Ma, W., Zhong, W., Feng, Z., Wang, H., Chen, Q., Peng, W., Feng, X., Qin, B., Liu, T.: A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions (Nov 2023). https://doi.org/10.48550/ARXIV.2311.05232
21. Huang, Y., Huang, J.: A survey on retrieval-augmented text generation for large language models (Apr 2024). https://doi.org/10.48550/ARXIV.2404.10981
22. Johnson, J., Douze, M., Jégou, H.: Billion-scale similarity search wip GPUs. IEEE Transactions on Big Data 7(3), 535–547 (2019)
23. Kamalloo, E., Thakur, N., Lassance, C., Ma, X., Yang, J.H., Lin, J.: Resources for brewing beir: Reproducible reference models and an official leaderboard (2023)
24. Khattab, O., Zaharia, M.: Colbert: Efficient and effective passage search via contextualized late interaction over bert (Apr 2020). https://doi.org/10.48550/ARXIV.2004.12832
25. Kwiatkowski, T., Palomaki, J., Redfield, O., Collins, M., Parikh, A., Alberti, C., Epstein, D., Polosukhin, I., Devlin, J., Lee, K., Toutanova, K., Jones, L., Kelcey, M., Chang, M.W., Dai, A.M., Uszkoreit, J., Le, Q., Petrov, S.: Natural questions: A benchmark for question answering research. Transactions of pe Association for Computational Linguistics 7, 453–466 (2019). https://doi.org/10.1162/tacl_a_00276, https://doi.org/10.1162/tacl_a_00276
26. Lanchantin, J., Toshniwal, S., Weston, J., Szlam, A., Sukhbaatar, S.: Learning to reason and memorize wip self-notes (May 2023). https://doi.org/10.48550/ARXIV.2305.00833
27. LangChain: Evaluating rag architectures on benchmark tasks (Nov 2023), https://langchain-ai.gipub.io/langchain-benchmarks/notebooks/retrieval/langchain_docs_qa.html
28. Leng, Q., Uhlenhup, K., Polyzotis, A.: Best Practices for LLM Evaluation of RAG Applications (Dec 2023), https://www.databricks.com/blog/LLM-auto-eval-best-practices-RAG
29. Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W.t., Rocktäschel, T., Riedel, S., Kiela, D.: Retrieval-augmented generation for knowledge-intensive NLP tasks. In: Proceedings of pe 34p International Conference on Neural Information Processing Systems. pp. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
9459–9474. NIPS’20, Curran Associates Inc., Red Hook, NY, USA (Dec 2020) | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# Evaluation of Retrieval-Augmented Generation: A Survey
|30.|Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W.t., Rocktäschel, T., Riedel, S., Kiela, D.: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. Tech. rep. (Apr 2021), http://arxiv.org/abs/2005.11401, arXiv:2005.11401 [cs] type: article|
|---|---|
|31.|Liang, X., Song, S., Niu, S., Li, Z., Xiong, F., Tang, B., Wy, Z., He, D., Cheng, P., Wang, Z., Deng, H.: Uhgeval: Benchmarking the hallucination of chinese large language models via unconstrained generation. arXiv preprint arXiv:2311.15296 (2023)|
|32.|Lin, C.Y.: ROUGE: A package for automatic evaluation of summaries. In: Text Summarization Branches Out. pp. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
74–81. Association for Computational Linguistics, Barcelona, Spain (Jul 2004), https://aclanthology.org/W04-1013|
|33.|Liu, Y., Huang, L., Li, S., Chen, S., Zhou, H., Meng, F., Zhou, J., Sun, X.: Recall: A benchmark for llms robustness against external counterfactual knowledge (Nov 2023). https://doi.org/10.48550/ARXIV.2311.08147|
|34.|Lyu, Y., Li, Z., Niu, S., Xiong, F., Tang, B., Wang, W., Wu, H., Liu, H., Xu, T., Chen, E., Luo, Y., Cheng, P., Deng, H., Wang, Z., Lu, Z.: Crud-rag: A comprehensive chinese benchmark for retrieval-augmented generation of large language models (Jan 2024). https://doi.org/10.48550/ARXIV.2401.17043|
|35.|Microsoft: Web Search API | Microsoft Bing, https://www.microsoft.com/en-us/bing/apis/bing-web-search-api|
|36.|Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: a method for automatic evaluation of machine translation. In: Isabelle, P., Charniak, E., Lin, D. (eds.) Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics. pp. 311–318. Association for Computational Linguistics, Philadelphia, Pennsylvania, USA (Jul 2002). https://doi.org/10.3115/1073083.1073135, https://aclanthology.org/P02-1040|
|37.|Petroni, F., Piktus, A., Fan, A., Lewis, P., Yazdani, M., De Cao, N., Thorne, J., Jernite, Y., Karpukhin, V., Maillard, J., Plachouras, V., Rocktäschel, T., Riedel, S.: KILT: a benchmark for knowledge intensive language tasks. In: Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. pp. 2523–2544. Association for Computational Linguistics, Online (Jun 2021). https://doi.org/10.18653/v1/2021.naacl-main.200, https://aclanthology.org/2021.naacl-main.200|
|38.|Ramos, J., et al.: Using tf-idf to determine word relevance in document queries. In: Proceedings of the first instructional conference on machine learning. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
vol. 242, pp. 29–48. Citeseer (2003)|
|39.|Robertson, S., Zaragoza, H., et al.: The probabilistic relevance framework: Bm25 and beyond. Foundations and Trends® in Information Retrieval 3(4), 333–389 (2009)|
|40.|Rosset, C., Chung, H.L., Qin, G., Chau, E.C., Feng, Z., Awadallah, A., Neville, J., Rao, N.: Researchy questions: A dataset of multi-perspective, decompositional questions for llm web agents (Feb 2024). https://doi.org/10.48550/ARXIV.2402.17896|
|41.|Saad-Falcon, J., Khattab, O., Potts, C., Zaharia, M.: Ares: An automated evaluation framework for retrieval-augmented generation systems (Nov 2023). https://doi.org/10.48550/ARXIV.2311.09476|
|42.|Shahabi, C., Kolahdouzan, M.R., Sharifzadeh, M.: A road network embedding technique for k-nearest neighbor search in moving object databases. In: Proceedings of the 10th ACM international symposium on advances in geographic information systems. pp. 94–100 (2002)|
|43.|Tang, Y., Yang, Y.: Multihop-rag: Benchmarking retrieval-augmented generation for multi-hop queries (Jan 2024). https://doi.org/10.48550/ARXIV.2401.15391|
|44.|Thorne, J., Vlachos, A., Christodoulopoulos, C., Mittal, A.: FEVER: a large-scale dataset for fact extraction and VERification. In: NAACL-HLT (2018)|
|45.|TruLens: TruLens (2023), https://www.trulens.org/trulens_eval/getting_started/quickstarts/quickstart/| | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |
# Hao Yu, Aoran Gan, Kai Zhang, Shiwei Tong†, Qi Liu, and Zhaofeng Liu
|46.|Wang, A., Pruksachatkun, Y., Nangia, N., Singh, A., Michael, J., Hill, F., Levy, O., Bowman, S.R.|SuperGLUE: A stickier benchmark for general-purpose language understanding systems. arXiv preprint 1905.00537 (2019)|
|---|---|---|
|47.|Wang, S., Khramtsova, E., Zhuang, S., Zuccon, G.|Feb4rag: Evaluating federated search in the context of retrieval augmented generation (Feb 2024). https://doi.org/10.48550/ARXIV.2402.11891|
|48.|Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q., Zhou, D.|Chain-of-thought prompting elicits reasoning in large language models (Jan 2022). https://doi.org/10.48550/ARXIV.2201.11903|
|49.|Xiong, G., Jin, Q., Lu, Z., Zhang, A.|Benchmarking retrieval-augmented generation for medicine (Feb 2024). https://doi.org/10.48550/ARXIV.2402.13178|
|50.|Xu, Z., Li, Y., Ding, R., Wang, X., Chen, B., Jiang, Y., Zheng, H.T., Lu, W., Xie, P., Huang, F.|Let llms take on the latest challenges! a chinese dynamic question answering benchmark (Feb 2024). https://doi.org/10.48550/ARXIV.2402.19248|
|51.|Yang, Z., Qi, P., Zhang, S., Bengio, Y., Cohen, W.W., Salakhutdinov, R., Manning, C.D.|HotpotQA: A dataset for diverse, explainable multi-hop question answering. In: Conference on Empirical Methods in Natural Language Processing (EMNLP) (2018)|
|52.|Yao, J.Y., Ning, K.P., Liu, Z.H., Ning, M.N., Yuan, L.|Llm lies: Hallucinations are not bugs, but features as adversarial examples. arXiv preprint arXiv:2310.01469 (2023)|
|53.|Zhang, Q., Chen, S., Xu, D., Cao, Q., Chen, X., Cohn, T., Fang, M.|A Survey for Efficient Open Domain Question Answering. In: Rogers, A., Boyd-Graber, J., Okazaki, N. | Provided the following context, I want to generate QA embedding pairs for each of my classes that are 3B and 7B LLM based on increasing complexity. For the 3B model, generate simple question-answer pairs, and for the 7B model, generate slightly more complex pairs. Please format the output in JSON as: {3B: {q1:, a1:}, {q2:, a2:}}, {7B: {q1:, a1:}, {q2:, a2:}} |