Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,86 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
datasets:
|
| 4 |
+
- hkust-nlp/deita-6k-v0
|
| 5 |
+
language:
|
| 6 |
+
- en
|
| 7 |
+
---
|
| 8 |
+
|
| 9 |
+
<img src="https://huggingface.co/datasets/hkust-nlp/deita-images/resolve/main/logo-final.png" alt="Deita banner" width="800" style="margin-left:'auto' margin-right:'auto' display:'block'"/>
|
| 10 |
+
|
| 11 |
+
# Model Card for Deita 7B V1.0
|
| 12 |
+
|
| 13 |
+
Deita is an open-sourced project designed to facilitate **Automatic Data Selection** for instruction tuning in Large Language Models (LLMs).
|
| 14 |
+
Deita 7B V1.0 is a fine-tuned + DPO version of Mistral-7B-v0.1 that was trained on 6k automatically selected lightweight, high-quality alignment SFT data: [Deita 6K V0](https://huggingface.co/datasets/hkust-nlp/deita-6k-v0) and 10K randomly sampled alignment preference data from Ultrafeedback.
|
| 15 |
+
|
| 16 |
+
## Model description
|
| 17 |
+
|
| 18 |
+
- **Model type:** Model trained on automatically selected lightweight, high-quality alignment SFT data and 10K randomly sampled alignment preference data.
|
| 19 |
+
- **Language(s) (NLP):** Primarily English
|
| 20 |
+
- **Finetuned from model:** Mistral-7B-v0.1
|
| 21 |
+
|
| 22 |
+
### Model Sources
|
| 23 |
+
|
| 24 |
+
- **Repository:** https://github.com/hkust-nlp/deita
|
| 25 |
+
- **Model Family:** Other models and the dataset are found in the [Deita collection](https://huggingface.co/collections/hkust-nlp/deita-6569c198c174808d94cf5bd4).
|
| 26 |
+
|
| 27 |
+
## Performance
|
| 28 |
+
| Model | Align | Data Size | MT-Bench | AlpacaEval(%) | OpenLLM (Avg.) |
|
| 29 |
+
|------------------------------------------------|-----------|------------|----------|---------------|----------------|
|
| 30 |
+
| **Proprietary Models** | | | | | |
|
| 31 |
+
| GPT-4-Turbo | ? | -- | 9.32 | 97.70 | -- |
|
| 32 |
+
| GPT-4 | SFT + PPO | -- | 8.99 | 95.03 | -- |
|
| 33 |
+
| Claude-2 | SFT + PPO | -- | 8.06 | 91.36 | -- |
|
| 34 |
+
| GPT-3.5-turbo | SFT + PPO | -- | 7.94 | 89.37 | -- |
|
| 35 |
+
| **Open-sourced Models based on Mistral-7B** | | | | | |
|
| 36 |
+
| Mistral-7B-Instruct-v0.1 | -- | -- | 6.84 | 69.65 | 60.45 |
|
| 37 |
+
| Zephyr-7B-sft | SFT | 200K SFT | 5.32 | 75.12 | 60.93 |
|
| 38 |
+
| Zephyr-7B-beta | SFT + DPO | 200K SFT + 60K DPO | 7.34 | 90.60 | 66.36 |
|
| 39 |
+
| OpenChat-3.5 | C-RLFT | >70K C-RLFT | 7.81 | 88.51 | -- |
|
| 40 |
+
| Starling-7B | C-RLFT + APA | >70K C-RLFT + 183K APA | 8.09 | 91.99 | -- |
|
| 41 |
+
| Random | SFT | 10K SFT | 5.89 | 56.90 | 61.72 |
|
| 42 |
+
| DEITA-7B-v1.0-sft | SFT | 6K SFT | 7.22 | 80.78 | 64.94 |
|
| 43 |
+
| DEITA-7B-v1.0-sft | SFT | 10K SFT | 7.32 | 81.67 | 64.00 |
|
| 44 |
+
| DEITA-7B-v1.0 | SFT + DPO | 6K SFT + 10K DPO | 7.55 | 90.06 | 69.86 |
|
| 45 |
+
|
| 46 |
+
## Input Format
|
| 47 |
+
|
| 48 |
+
The model is trained using the [vicuna_v1.1 template](https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py)
|
| 49 |
+
|
| 50 |
+
### SFT Format
|
| 51 |
+
|
| 52 |
+
```
|
| 53 |
+
A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: Hello! ASSISTANT: Hi!</s>USER: How are you? ASSISTANT:
|
| 54 |
+
```
|
| 55 |
+
|
| 56 |
+
### DPO Format
|
| 57 |
+
|
| 58 |
+
```
|
| 59 |
+
A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <prompt> ASSISTANT: <answer></s>
|
| 60 |
+
```
|
| 61 |
+
|
| 62 |
+
where \<answer\> can be a chosen answer or a rejected answer.
|
| 63 |
+
|
| 64 |
+
### Training hyperparameters
|
| 65 |
+
|
| 66 |
+
The following hyperparameters were used during training:
|
| 67 |
+
- learning_rate: 2e-05
|
| 68 |
+
- train_batch_size: 1
|
| 69 |
+
- eval_batch_size: 1
|
| 70 |
+
- seed: 42
|
| 71 |
+
- distributed_type: multi-GPU
|
| 72 |
+
- num_devices: 4
|
| 73 |
+
- gradient_accumulation_steps: 128
|
| 74 |
+
- total_train_batch_size: 512
|
| 75 |
+
- total_eval_batch_size: 4
|
| 76 |
+
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
|
| 77 |
+
- lr_scheduler_type: cosine
|
| 78 |
+
- lr_scheduler_warmup_ratio: 0.1
|
| 79 |
+
- num_epochs: 6.0
|
| 80 |
+
|
| 81 |
+
### Framework versions
|
| 82 |
+
|
| 83 |
+
- Transformers 4.34.1
|
| 84 |
+
- Pytorch 2.1.0+cu121
|
| 85 |
+
- Datasets 2.14.6
|
| 86 |
+
- Tokenizers 0.14.1
|