This model was trained on our MetamathFewshot dataset, as well as the Vicuna dataset and the OrcaChat dataset.
It has been finetuned from base Mistral 7B
Usage
This model uses a specific prompt format which is encoded as a chat template. To apply this, you can use the tokenizer.apply_chat_template() method of the attached tokenizer:
messages = [
{"role": "user", "content": "What is the capital of Spain?"},
{"role": "assistant", "content": "The capital of Spain is Madrid."}
]
gen_input = tokenizer.apply_chat_template(message, return_tensors="pt")
model.generate(**gen_input)
Evaluation Results
HuggingFace Leaderboard
| Average | ARC | HellaSwag | MMLU | TruthfulQA | Winogrande | GSM8K |
|---|---|---|---|---|---|---|
| 67.33 | 59.64 | 81.82 | 61.69 | 53.23 | 78.45 | 69.14 |
For comparison the GSM8K score for the original metamath/MetaMath-Mistral-7B was 68.84 and average score was 65.78.
MT-Bench
| Turn 1 | Turn 2 | Average |
|---|---|---|
| 6.90 | 6.52 | 6.71 |
Training Details
Instruction tuned with the following parameters:
- LORA, Rank 8, Alpha 16, Dropout 0.05, all modules (QKV and MLP)
- 3 epochs
- Micro Batch Size 32 over 4xH100, gradient accumulation steps = 1
- AdamW with learning rate 5e-5
Bias, Risks, and Limitations
The model has not been evaluated for safety and is only intended for research and experiments.