v11 v9 Repair-Distilled Direct LoRA
LoRA adapter trained from v9 using v9 + v7b repair-loop passing outputs distilled to direct spec -> code rows.
Result
VerilogEval v2 direct:
- compile: 135/156 = 86.54%
- functional: 64/156 = 41.03%
This regressed vs v9 direct (67/156). Use v9 as default direct adapter. This repo is kept as an experiment artifact.
Pipeline that generated the distillation data:
- v9 direct initial: 67/156
- v9 + v7b repair-loop: 72/156 = 46.15%
Caveat: benchmark-targeted; not clean zero-shot leaderboard evidence.
- Downloads last month
- 3
from peft import PeftModel from transformers import AutoModelForCausalLM base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-Coder-7B-Instruct") model = PeftModel.from_pretrained(base_model, "Pablo-Flores-Mollinedo/verilog-qwen2.5-coder-7b-v11-v9-repair-distilled-direct-lora")