Asif Ahmad commited on
Commit
9f4882b
1 Parent(s): 114674f

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +109 -0
README.md ADDED
@@ -0,0 +1,109 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: llama2
3
+ datasets:
4
+ - togethercomputer/RedPajama-Data-1T
5
+ - togethercomputer/RedPajama-Data-Instruct
6
+ - EleutherAI/pile
7
+ - togethercomputer/Long-Data-Collections
8
+ language:
9
+ - en
10
+ library_name: transformers
11
+ ---
12
+
13
+ # LLaMA-2-7B-32K
14
+
15
+ ## Model Description
16
+
17
+ LLaMA-2-7B-32K is an open-source, long context language model developed by Together, fine-tuned from Meta's original Llama-2 7B model.
18
+ This model represents our efforts to contribute to the rapid progress of the open-source ecosystem for large language models.
19
+ The model has been extended to a context length of 32K with position interpolation,
20
+ allowing applications on multi-document QA, long text summarization, etc.
21
+
22
+ ## What's new?
23
+
24
+ This model introduces several improvements and new features:
25
+
26
+ 1. **Extended Context:** The model has been trained to handle context lengths up to 32K, which is a significant improvement over the previous versions.
27
+
28
+ 2. **Pre-training and Instruction Tuning:** We have shared our data recipe, which consists of a mixture of pre-training and instruction tuning data.
29
+
30
+ 3. **Fine-tuning Examples:** We provide examples of how to fine-tune the model for specific applications, including book summarization and long context question and answering.
31
+
32
+ 4. **Software Support:** We have updated both the inference and training stack to allow efficient inference and fine-tuning for 32K context.
33
+
34
+ ## Model Architecture
35
+
36
+ The model follows the architecture of Llama-2-7B and extends it to handle a longer context. It leverages the recently released FlashAttention-2 and a range of other optimizations to improve the speed and efficiency of inference and training.
37
+
38
+ ## Training and Fine-tuning
39
+
40
+ The model has been trained using a mixture of pre-training and instruction tuning data.
41
+ - In the first training phase of continued pre-training, our data mixture contains 25% RedPajama Book, 25% RedPajama ArXiv (including abstracts), 25% other data from RedPajama, and 25% from the UL2 Oscar Data, which is a part of OIG (Open-Instruction-Generalist), asking the model to fill in missing chunks, or complete the text.
42
+ To enhance the long-context ability, we exclude data shorter than 2K word. The inclusion of UL2 Oscar Data is effective in compelling the model to read and utilize long-range context.
43
+ - We then fine-tune the model to focus on its few shot capacity under long context, including 20% Natural Instructions (NI), 20% Public Pool of Prompts (P3), 20% the Pile. We decontaminated all data against HELM core scenarios . We teach the model to leverage the in-context examples by packing examples into one 32K-token sequence. To maintain the knowledge learned from the first piece of data, we incorporate 20% RedPajama-Data Book and 20% RedPajama-Data ArXiv.
44
+
45
+ Next, we provide examples of how to fine-tune the model for specific applications.
46
+ The example datasets are placed in [togethercomputer/Long-Data-Collections](https://huggingface.co/datasets/togethercomputer/Long-Data-Collections)
47
+ You can use the [OpenChatKit](https://github.com/togethercomputer/OpenChatKit) to fine-tune your own 32K model over LLaMA-2-7B-32K.
48
+ Please refer to [OpenChatKit](https://github.com/togethercomputer/OpenChatKit) for step-by-step illustrations.
49
+
50
+ 1. Long Context QA.
51
+
52
+ We take as an example the multi-document question answering task from the paper “Lost in the Middle: How Language Models Use Long Contexts”. The input for the model consists of (i) a question that requires an answer and (ii) k documents, which are passages extracted from Wikipedia. Notably, only one of these documents contains the answer to the question, while the remaining k − 1 documents, termed as "distractor" documents, do not. To successfully perform this task, the model must identify and utilize the document containing the answer from its input context.
53
+
54
+ With OCK, simply run the following command to fine-tune:
55
+ ```
56
+ bash training/finetune_llama-2-7b-32k-mqa.sh
57
+ ```
58
+
59
+ 2. Summarization.
60
+
61
+ Another example is BookSum, a unique dataset designed to address the challenges of long-form narrative summarization. This dataset features source documents from the literature domain, including novels, plays, and stories, and offers human-written, highly abstractive summaries. We here focus on chapter-level data. BookSum poses a unique set of challenges, necessitating that the model comprehensively read through each chapter.
62
+
63
+ With OCK, simply run the following command to fine-tune:
64
+ ```
65
+ bash training/finetune_llama-2-7b-32k-booksum.sh
66
+ ```
67
+
68
+
69
+ ## Inference
70
+
71
+ You can use the [Together API](https://together.ai/blog/api-announcement) to try out LLaMA-2-7B-32K for inference.
72
+ The updated inference stack allows for efficient inference.
73
+
74
+ To run the model locally, we strongly recommend to install Flash Attention V2, which is necessary to obtain the best performance:
75
+ ```
76
+ # Please update the path of `CUDA_HOME`
77
+ export CUDA_HOME=/usr/local/cuda-11.8
78
+ pip install transformers==4.31.0
79
+ pip install sentencepiece
80
+ pip install ninja
81
+ pip install flash-attn --no-build-isolation
82
+ pip install git+https://github.com/HazyResearch/flash-attention.git#subdirectory=csrc/rotary
83
+ ```
84
+
85
+ You can use this model directly from the Hugging Face Model Hub or fine-tune it on your own data using the OpenChatKit.
86
+
87
+ ```python
88
+ from transformers import AutoTokenizer, AutoModelForCausalLM
89
+
90
+ tokenizer = AutoTokenizer.from_pretrained("togethercomputer/LLaMA-2-7B-32K")
91
+ model = AutoModelForCausalLM.from_pretrained("togethercomputer/LLaMA-2-7B-32K", trust_remote_code=True, torch_dtype=torch.float16)
92
+
93
+ input_context = "Your text here"
94
+ input_ids = tokenizer.encode(input_context, return_tensors="pt")
95
+ output = model.generate(input_ids, max_length=128, temperature=0.7)
96
+ output_text = tokenizer.decode(output[0], skip_special_tokens=True)
97
+ print(output_text)
98
+ ```
99
+
100
+ Alternatively, you can set `trust_remote_code=False` if you prefer not to use flash attention.
101
+
102
+
103
+ ## Limitations and Bias
104
+
105
+ As with all language models, LLaMA-2-7B-32K may generate incorrect or biased content. It's important to keep this in mind when using the model.
106
+
107
+ ## Community
108
+
109
+ Join us on [Together Discord](https://discord.gg/6ZVDU8tTD4)