Evaluation (Standardized)
Goal
Make results comparable across community fine-tunes.
Recommended metrics
- HumanEval pass@1
- Optional: MBPP pass@1
Suggested tool
Use lm-evaluation-harness (or your preferred harness) to run HumanEval and report settings:
- base model
- training recipe (full / LoRA / QLoRA)
- sequence length
- epochs
- hardware