luminus-coder-7b-sft-adapter
LoRA adapter (r=32, attention projections) for Qwen2.5-Coder-7B-Instruct, fine-tuned on verified code solutions. Load with PeftModel.from_pretrained(base_model, this_repo).
Eval (held-out set of 325 coding problems, k=4, unit-test verified)
| pass@1 | pass@4 | |
|---|---|---|
| base (Qwen2.5-Coder-7B-Instruct) | 0.112 | 0.169 |
| + this adapter (original run) | 0.152 | 0.246 |
| + this adapter (reproduction run) | 0.144 | 0.225 |
Trained model solves 36% more held-out problems (pass@1) and 46% more within 4 tries (pass@4) than its base. The reproduction run (same data, same config, retrained from scratch) confirms the gain within run-to-run variance (~1pp).
- Downloads last month
- 27
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support