Text Generation
Transformers
PyTorch
Safetensors
English
llama
conversational
text-generation-inference
Inference Endpoints
hamishivi commited on
Commit
bcfffa1
1 Parent(s): b930928

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +136 -0
README.md ADDED
@@ -0,0 +1,136 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ model-index:
3
+ - name: tulu-2-dpo-13b
4
+ results: []
5
+ datasets:
6
+ - HuggingFaceH4/ultrafeedback_binarized
7
+ - allenai/tulu-v2-sft-mixture
8
+ language:
9
+ - en
10
+ base_model: meta-llama/Llama-2-13b-hf
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 Tulu V2 DPO 13B
18
+
19
+ Tulu is a series of language models that are trained to act as helpful assistants.
20
+ Tulu V2 DPO 70B, and is a fine-tuned version of Llama 2 that was trained on on a mix of publicly available, synthetic and human datasets using [Direct Preference Optimization (DPO)](https://arxiv.org/abs/2305.18290).
21
+ This model is a strong alternative to Llama 2 13b Chat.
22
+
23
+
24
+ ## Model description
25
+
26
+ - **Model type:** The flagship model of a suite of instruction and RLHF tuned chat models on a mix of publicly available, synthetic and human-created datasets.
27
+ - **Language(s) (NLP):** Primarily English
28
+ - **License:** [AI2 ImpACT](https://allenai.org/impact-license) Low-risk license.
29
+ - **Finetuned from model:** [meta-llama/Llama-2-70b-hf](https://huggingface.co/meta-llama/Llama-2-70b-hf)
30
+
31
+ ### Model Sources
32
+
33
+ - **Repository:** https://github.com/allenai/https://github.com/allenai/open-instruct
34
+ - **DPO Recipe:** The DPO recipe is from the [Zephyr Beta](https://huggingface.co/HuggingFaceH4/zephyr-7b-beta) model
35
+ - **Model Family:** Other models and the dataset are found in the [Tulu V2 collection](https://huggingface.co/collections/allenai/tulu-v2-suite-6551b56e743e6349aab45101).
36
+
37
+ ## Performance
38
+
39
+ | Model | Size | Alignment | MT-Bench (score) | AlpacaEval (win rate %) |
40
+ |-------------|-----|----|---------------|--------------|
41
+ | **Tulu-v2-7b** 🐪 | **7B** | **dDPO** | **6.30** | **73.9** |
42
+ | **Tulu-v2-dpo-7b** 🐪 | **7B** | **dDPO** | **6.27** | **85.1** |
43
+ | StableLM-Tuned-α | 7B| dSFT |2.75| -|
44
+ | MPT-Chat | 7B |dSFT |5.42| -|
45
+ | Xwin-LMv0.1 | 7B| dPPO| 6.19| 87.83|
46
+ | Mistral-Instructv0.1 | 7B| - | 6.84 |-|
47
+ | Zephyr-7b-α |7B| dDPO| 6.88| -|
48
+ | Zephyr-7b-β 🪁 | 7B | dDPO | 7.34 | 90.60 |
49
+ | **Tulu-v2-13b** 🐪 | **13B** | **dDPO** | **6.70** | **78.9** |
50
+ | **Tulu-v2-dpo-13b** 🐪 | **13B** | **dDPO** | **7.00** | **89.5** |
51
+ | Falcon-Instruct | 40B |dSFT |5.17 |45.71|
52
+ | Guanaco | 65B | SFT |6.41| 71.80|
53
+ | Llama2-Chat | 70B |RLHF |6.86| 92.66|
54
+ | Vicuna v1.3 | 33B |dSFT |7.12 |88.99|
55
+ | WizardLM v1.0 | 70B |dSFT |7.71 |-|
56
+ | Xwin-LM v0.1 | 70B |dPPO |- |95.57|
57
+ | **Tulu-v2-70b** 🐪 | **70B** | **dDPO** | **7.49** | **86.6** |
58
+ | **Tulu-v2-dpo-70b** 🐪 | **70B** | **dDPO** | **7.89** | **95.1** |
59
+ | GPT-3.5-turbo | - |RLHF |7.94 |89.37|
60
+ | Claude 2 | - |RLHF |8.06| 91.36|
61
+ | GPT-4 | -| RLHF |8.99| 95.28|
62
+
63
+
64
+ ## Intended uses & limitations
65
+
66
+ The model was initially fine-tuned on a filtered and preprocessed of 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.
67
+ 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.
68
+
69
+
70
+ <!-- You can find the datasets used for training Tulu V2 [here]()
71
+
72
+ Here's how you can run the model using the `pipeline()` function from 🤗 Transformers:
73
+
74
+ ```python
75
+ # Install transformers from source - only needed for versions <= v4.34
76
+ # pip install git+https://github.com/huggingface/transformers.git
77
+ # pip install accelerate
78
+
79
+ import torch
80
+ from transformers import pipeline
81
+
82
+ pipe = pipeline("text-generation", model="HuggingFaceH4/tulu-2-dpo-70b", torch_dtype=torch.bfloat16, device_map="auto")
83
+
84
+ # We use the tokenizer's chat template to format each message - see https://huggingface.co/docs/transformers/main/en/chat_templating
85
+ messages = [
86
+ {
87
+ "role": "system",
88
+ "content": "You are a friendly chatbot who always responds in the style of a pirate",
89
+ },
90
+ {"role": "user", "content": "How many helicopters can a human eat in one sitting?"},
91
+ ]
92
+ prompt = pipe.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
93
+ outputs = pipe(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
94
+ print(outputs[0]["generated_text"])
95
+ # <|system|>
96
+ # You are a friendly chatbot who always responds in the style of a pirate.</s>
97
+ # <|user|>
98
+ # How many helicopters can a human eat in one sitting?</s>
99
+ # <|assistant|>
100
+ # Ah, me hearty matey! But yer question be a puzzler! A human cannot eat a helicopter in one sitting, as helicopters are not edible. They be made of metal, plastic, and other materials, not food!
101
+ ```-->
102
+
103
+ ## Bias, Risks, and Limitations
104
+
105
+ <!-- This section is meant to convey both technical and sociotechnical limitations. -->
106
+
107
+ 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).
108
+ It is also unknown what the size and composition of the corpus was used to train the base Llama 2 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.
109
+
110
+
111
+ ### Training hyperparameters
112
+
113
+ The following hyperparameters were used during DPO training:
114
+ - learning_rate: 5e-07
115
+ - total_train_batch_size: 32
116
+ - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
117
+ - lr_scheduler_type: linear
118
+ - lr_scheduler_warmup_ratio: 0.1
119
+ - num_epochs: 3.0
120
+
121
+
122
+ ## Citation
123
+
124
+ If you find Tulu 2 is useful in your work, please cite it with:
125
+
126
+ ```
127
+ @misc{ivison2023changing,
128
+ title={Camels in a Changing Climate: Enhancing LM Adaptation with Tulu 2},
129
+ 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},
130
+ year={2023},
131
+ archivePrefix={arXiv},
132
+ primaryClass={cs.CL}
133
+ }
134
+ ```
135
+
136
+ *Model card adapted from [Zephyr Beta](https://huggingface.co/HuggingFaceH4/zephyr-7b-beta/blob/main/README.md)*