README.md

metadata

model-index:
  - name: MetaMath-LoRA-LLaMA-7B
    results:
      - task:
          type: text-generation
        dataset:
          name: meta-math/MetaMathQA
          type: meta-math/MetaMathQA
        metrics:
          - name: Accuracy (zero-shot)
            type: Accuracy (zero-shot)
            value: 0.62
            verified: true
        source:
          name: Arithmetic Reasoning on GSM8K
          url: https://paperswithcode.com/sota/arithmetic-reasoning-on-gsm8k
license: mit

Fine-tune LLaMA 2 (7B) with LoRA on meta-math/MetaMathQA

Fine-tune for three epochs

Result:

Reload the saved adapter: Invalid output length: 6, Testing length: 1319, Accuracy: 0.62

Comparison

The official report accuracy is 0.665 by fine-tuning the whole LLaMA 2 7B model for 3 epochs.

Note: The LoRA adapter is being used for future research purposes.