Text Generation
Transformers
Safetensors
English
llama
conversational
text-generation-inference
Inference Endpoints
hamishivi commited on
Commit
c5f3b79
1 Parent(s): 63cb783

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +100 -0
README.md ADDED
@@ -0,0 +1,100 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ model-index:
3
+ - name: llama-3-tulu-2-dpo-70b
4
+ results: []
5
+ datasets:
6
+ - allenai/tulu-v2-sft-mixture
7
+ - argilla/ultrafeedback-binarized-preferences-cleaned
8
+ language:
9
+ - en
10
+ base_model: allenai/llama-3-tulu-2-70b
11
+ ---
12
+
13
+
14
+ <img src="https://huggingface.co/datasets/allenai/blog-images/resolve/main/tulu-v2/Tulu%20V2%20banner.png" alt="TuluV2 banner" width="800" style="margin-left:'auto' margin-right:'auto' display:'block'"/>
15
+
16
+
17
+ # Model Card for Llama 3 Tulu V2+DPO 70B
18
+
19
+ Tulu is a series of language models that are trained to act as helpful assistants.
20
+ Llama 3 Tulu V2 70B is a fine-tuned version of Llama 3 that was trained on a mix of publicly available, synthetic and human datasets.
21
+ It was then further trained using [DPO]((https://arxiv.org/abs/2305.18290) on the [UltraFeedback](https://huggingface.co/datasets/argilla/ultrafeedback-binarized-preferences-cleaned) dataset.
22
+
23
+ For more details on the training mixture, read the paper: [Camels in a Changing Climate: Enhancing LM Adaptation with Tulu 2
24
+ ](https://arxiv.org/abs/2311.10702).
25
+
26
+
27
+ ## Model description
28
+
29
+ - **Model type:** A model trained on a mix of publicly available, synthetic and human-created datasets.
30
+ - **Language(s) (NLP):** Primarily English
31
+ - **License:** Apache 2.0
32
+ - **Finetuned from model:** [meta-llama/Meta-Llama-3-70B](https://huggingface.co/meta-llama/Meta-Llama-3-70B)
33
+
34
+ ### Model Sources
35
+
36
+ - **Repository:** https://github.com/allenai/open-instruct
37
+ - **Model Family:** Other models and the dataset are found in the [Tulu V2 collection](https://huggingface.co/collections/allenai/tulu-v2-suite-6551b56e743e6349aab45101).
38
+
39
+ ## Performance
40
+
41
+ | Model | MMLU 5-shot | GSM8k 8-shot cot | BBH 3-shot cot | TydiQA 1-shot Gold Passage | Codex HumanEval Pass@10 |AlpacaEval 1 | AlpacaEval 2 LC | TruthfulQA %Info+True | IFEval loose acc | XSTest safe but ref. | XSTest unsafe but follow | Average |
42
+ |-|-|-|-|-|-|-|-|-|-|-|-|-|
43
+ | Llama 3 70b base | 0.790 | 0.840 | 0.801 | 73.35 | 0.745 | - | - | 0.469 | 0.163 | 0.256 | 0.330 | 65.60 |
44
+ | Llama 3 70b instruct | 0.786 | 0.930 | 0.801 | 59.21 | 0.908 | 96.71 | 39.99 | 0.701 | 0.828 | 0.060 | 0.140 | 79.22 |
45
+ | Llama 3 Tulu 2 70b | 0.752 | 0.845 | 0.779 | 69.798 | 0.861 | 86.007 | 17.51 | 0.646 | 0.591 | 0.108 | 0.130 | 73.01 |
46
+ | Llama 3 Tulu 2+DPO 70b | 0.754 | 0.860 | 0.785 | 23.443 | 0.878 | 96.65 | 27.34 | 0.780 | 0.643 | 0.080 | 0.140 | 71.60 |
47
+
48
+
49
+ ## Input Format
50
+
51
+ The model is trained to use the following format (note the newlines):
52
+ ```
53
+ <|user|>
54
+ Your message here!
55
+ <|assistant|>
56
+ ```
57
+
58
+ For best results, format all inputs in this manner. **Make sure to include a newline after `<|assistant|>`, this can affect generation quality quite a bit.**
59
+
60
+
61
+ ## Intended uses & limitations
62
+
63
+ The model was initially fine-tuned on the [Tulu V2 mix dataset](https://huggingface.co/datasets/allenai/tulu-v2-sft-mixture), which contains a diverse range of human created instructions and synthetic dialogues generated primarily by other LLMs.
64
+ We then further aligned the model with a [Jax DPO trainer](https://github.com/hamishivi/EasyLM/blob/main/EasyLM/models/llama/llama_train_dpo.py) built on [EasyLM](https://github.com/young-geng/EasyLM) on the [openbmb/UltraFeedback](https://huggingface.co/datasets/openbmb/UltraFeedback) dataset, which contains 64k prompts and model completions that are ranked by GPT-4.
65
+
66
+
67
+ ## Bias, Risks, and Limitations
68
+
69
+
70
+ The Tulu models have not been aligned to generate safe completions within the RLHF phase or deployed with in-the-loop filtering of responses like ChatGPT, so the model can produce problematic outputs (especially when prompted to do so).
71
+ It is also unknown what the size and composition of the corpus was used to train the base Llama 3 models, however it is likely to have included a mix of Web data and technical sources like books and code. See the [Falcon 180B model card](https://huggingface.co/tiiuae/falcon-180B#training-data) for an example of this.
72
+
73
+
74
+ ### DPO Training hyperparameters
75
+
76
+ The following hyperparameters were used during DPO training:
77
+ - learning_rate: 5e-7
78
+ - total_train_batch_size: 32
79
+ - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
80
+ - lr_scheduler_type: linear
81
+ - lr_scheduler_warmup_ratio: 0.1
82
+ - num_epochs: 3.0
83
+
84
+
85
+ ## Citation
86
+
87
+ If you find Tulu 2 is useful in your work, please cite it with:
88
+
89
+ ```
90
+ @misc{ivison2023camels,
91
+ title={Camels in a Changing Climate: Enhancing LM Adaptation with Tulu 2},
92
+ author={Hamish Ivison and Yizhong Wang and Valentina Pyatkin and Nathan Lambert and Matthew Peters and Pradeep Dasigi and Joel Jang and David Wadden and Noah A. Smith and Iz Beltagy and Hannaneh Hajishirzi},
93
+ year={2023},
94
+ eprint={2311.10702},
95
+ archivePrefix={arXiv},
96
+ primaryClass={cs.CL}
97
+ }
98
+ ```
99
+
100
+ *Model card adapted from [Zephyr Beta](https://huggingface.co/HuggingFaceH4/zephyr-7b-beta/blob/main/README.md)*