Pablo-Flores-Mollinedo/verilog-qwen2.5-coder-7b-v9-k5-v7b-repair-pipeline

Verilog Qwen2.5-Coder 7B v9 k=5 + v7b Repair Pipeline

This repository documents the best current VerilogEval v2 inference pipeline from the project.

Components

Generator adapter:

• —Pablo-Flores-Mollinedo/verilog-qwen2.5-coder-7b-v9-auto-distilled-direct-lora

Repair adapter:

• —Pablo-Flores-Mollinedo/verilog-qwen2.5-coder-7b-v7b-repair-loop-lora

Base model:

• —Qwen/Qwen2.5-Coder-7B-Instruct

Procedure

1. 1.Generate up to 5 direct candidates with the v9 adapter.
2. 2.Compile/simulate candidates.
3. 3.Keep a passing direct candidate when available.
4. 4.For remaining failing tasks, run v7b repair using compiler/simulator logs.
5. 5.Compile/simulate repaired RTL.

Result

VerilogEval v2 spec-to-RTL:

{
  "total": 156,
  "k_passed": 72,
  "repair_passed": 4,
  "final_unique_rows": 76,
  "pass_pct": 48.72
}

Best direct adapter for comparison:

v9 direct: 67/156 = 42.95%

Caveat

This is an inference pipeline, not a single adapter. It is benchmark-targeted and should not be reported as a clean zero-shot leaderboard claim.