Pablo-Flores-Mollinedo/verilog-qwen2.5-coder-7b-v6-error-driven-repair-lora

Verilog Qwen2.5-Coder 7B v6 Error-Driven Repair LoRA

LoRA adapter for Qwen/Qwen2.5-Coder-7B-Instruct trained from the v4.1 Verilog adapter using error-driven repair examples.

Important caveat

This adapter uses VerilogEval-derived repair data. VerilogEval v2 numbers are therefore benchmark-targeted repair results, not clean zero-shot leaderboard placement.

For clean/public production use, v4.1 remains the recommended adapter unless you specifically want VerilogEval-targeted repair behavior.

Training format

original spec
+ incorrect generated code
+ compile/simulation failure hint
-> corrected verified TopModule

VerilogEval v2 direct score

Spec-to-RTL, 156 tasks, n=1, temperature=0, top_p=0.01.

Adapter	Compile	Functional
v4.1	82.05%	36.54%
v5b	85.26%	36.54%
v6	84.62%	37.82%

v6 vs v4.1:

compile:    +2.56 points
functional: +1.28 points

Dataset

data/v6_error_driven_repair.jsonl
508 examples
96 validated repairs

Load

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "Qwen/Qwen2.5-Coder-7B-Instruct"
adapter = "Pablo-Flores-Mollinedo/verilog-qwen2.5-coder-7b-v6-error-driven-repair-lora"

tokenizer = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, device_map="auto")
model = PeftModel.from_pretrained(model, adapter)