lewtun HF staff commited on
Commit
4372f6c
1 Parent(s): c0b9e47

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +46 -17
README.md CHANGED
@@ -1,28 +1,24 @@
1
  ---
2
- base_model: HuggingFaceH4/mistral-7b-ift
3
  tags:
4
  - generated_from_trainer
5
  model-index:
6
- - name: mistral-7b-dpo-v0.4
7
  results: []
 
 
 
 
 
 
8
  ---
9
 
10
  <!-- This model card has been generated automatically according to the information the Trainer had access to. You
11
  should probably proofread and complete it, then remove this comment. -->
12
 
13
- # mistral-7b-dpo-v0.4
 
 
14
 
15
- This model is a fine-tuned version of [HuggingFaceH4/mistral-7b-ift](https://huggingface.co/HuggingFaceH4/mistral-7b-ift) on the HuggingFaceH4/ultrafeedback dataset.
16
- It achieves the following results on the evaluation set:
17
- - Loss: 0.4605
18
- - Rewards/chosen: -0.5053
19
- - Rewards/rejected: -1.8752
20
- - Rewards/accuracies: 0.7812
21
- - Rewards/margins: 1.3699
22
- - Logps/rejected: -327.4286
23
- - Logps/chosen: -297.1040
24
- - Logits/rejected: -2.7153
25
- - Logits/chosen: -2.7447
26
 
27
  ## Model description
28
 
@@ -30,11 +26,44 @@ More information needed
30
 
31
  ## Intended uses & limitations
32
 
33
- More information needed
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
34
 
35
  ## Training and evaluation data
36
 
37
- More information needed
 
 
 
 
 
 
 
 
 
 
38
 
39
  ## Training procedure
40
 
@@ -84,4 +113,4 @@ The following hyperparameters were used during training:
84
  - Transformers 4.34.0
85
  - Pytorch 2.0.1+cu118
86
  - Datasets 2.12.0
87
- - Tokenizers 0.14.0
 
1
  ---
 
2
  tags:
3
  - generated_from_trainer
4
  model-index:
5
+ - name: zephyr-7b-alpha
6
  results: []
7
+ license: cc-by-nc-4.0
8
+ datasets:
9
+ - stingning/ultrachat
10
+ - openbmb/UltraFeedback
11
+ language:
12
+ - en
13
  ---
14
 
15
  <!-- This model card has been generated automatically according to the information the Trainer had access to. You
16
  should probably proofread and complete it, then remove this comment. -->
17
 
18
+ # Zephyr 7B Alpha
19
+
20
+ Zephyr is a series of language models that are trained to act as helpful assistants. Zephyr-7B-α is the first model in the series, and is a fine-tuned version of [mistralai/Mistral-7B-v0.1](https://huggingface.co/mistralai/Mistral-7B-v0.1) that was trained on on a mix of publicly available, synthetic datasets using [Direct Preference Optimization (DPO)](https://arxiv.org/abs/2305.18290). We found that removing the in-built alignment of these datasets boosted performance on [MT Bench](https://huggingface.co/spaces/lmsys/mt-bench) and made the model more helpful. However, this means that model is likely to generate problematic text when prompted to do so and should only be used for educational and research purposes.
21
 
 
 
 
 
 
 
 
 
 
 
 
22
 
23
  ## Model description
24
 
 
26
 
27
  ## Intended uses & limitations
28
 
29
+ The model was initially fine-tuned on a variant of the [`UltraChat`](https://huggingface.co/datasets/stingning/ultrachat) dataset, which contains a diverse range of synthetic dialogues generated by ChatGPT. We then further aligned the model with [🤗 TRL's](https://github.com/huggingface/trl) `DPOTrainer` on the [openbmb/UltraFeedback](https://huggingface.co/datasets/openbmb/UltraFeedback) dataset, which contain 64k prompts and model completions that are ranked by GPT-4. As a result, the model can be used for chat and you can check out our [demo](https://huggingface.co/spaces/HuggingFaceH4/zephyr-playground) to test its capabilities.
30
+
31
+ Here's how you can run the model using the `pipeline()` function from 🤗 Transformers:
32
+
33
+ ```python
34
+ import torch
35
+ from transformers import pipeline
36
+
37
+ pipe = pipeline("text-generation", model="HuggingFaceH4/zephyr-7b-alpha", torch_dtype=torch.bfloat16, device_map="auto")
38
+
39
+ # We use a variant of ChatML to format each message
40
+ prompt_template = "<|system|>\n</s>\n<|user|>\n{query}</s>\n<|assistant|>\n"
41
+ prompt = prompt_template.format(query="How many helicopters can a human eat in one sitting?")
42
+ outputs = pipe(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
43
+ # Zero. Humans cannot consume or digest solid objects like helicopters, including their components such as rotor blades and engines. A human's diet is limited to food that they can swallow and break down through the process of digestion. Eating a helicopter would be physically impossible and could potentially cause serious harm if attempted.
44
+ ```
45
+
46
+ ## Bias, Risks, and Limitations
47
+
48
+ <!-- This section is meant to convey both technical and sociotechnical limitations. -->
49
+
50
+ Zephyr-7B-α has not been aligned to human preferences with techniques like RLHF or deployed with in-the-loop filtering of responses like ChatGPT, so the model can produce problematic outputs (especially when prompted to do so).
51
+ It is also unknown what the size and composition of the corpus was used to train the base model (`mistralai/Mistral-7B-v0.1`), 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.
52
+
53
 
54
  ## Training and evaluation data
55
 
56
+ Zephyr 7B Alpha achieves the following results on the evaluation set:
57
+
58
+ - Loss: 0.4605
59
+ - Rewards/chosen: -0.5053
60
+ - Rewards/rejected: -1.8752
61
+ - Rewards/accuracies: 0.7812
62
+ - Rewards/margins: 1.3699
63
+ - Logps/rejected: -327.4286
64
+ - Logps/chosen: -297.1040
65
+ - Logits/rejected: -2.7153
66
+ - Logits/chosen: -2.7447
67
 
68
  ## Training procedure
69
 
 
113
  - Transformers 4.34.0
114
  - Pytorch 2.0.1+cu118
115
  - Datasets 2.12.0
116
+ - Tokenizers 0.14.0