Mistral-7B-v0.1 fine-tuned for science multiple-choice QA
- Stage 1: LoRA SFT on ARC-Challenge + ARC-Easy train (answer tokens only,
Question: ...\nAnswer: ...format) - Stage 2: LoRA DPO on OpenBookQA train (chosen = correct option, rejected = distractors, beta = 0.1)
- LoRA adapters merged into fp16 weights
| ARC-Challenge (lm-eval, 25-shot) | acc | acc_norm |
|---|---|---|
| Mistral-7B-v0.1 | 57.00 | 61.43 |
| This model | 71.67 | 74.74 |
- Downloads last month
- 2
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for wonjunes/mistral-7b-arc
Base model
mistralai/Mistral-7B-v0.1