Pablo-Flores-Mollinedo/verilog-rtl-lora-adapter-suite

Verilog RTL LoRA Adapter Suite (TCAD 2026 Draft Artifacts)

This repository collects the main PEFT LoRA adapters used in the adapter-centric Verilog RTL generation study. The original local adapter directories were copied and renamed with descriptive artifact names. No original local adapter directory was deleted or modified.

Adapter mapping

Descriptive artifact name	Paper alias	Base/start	Intended role
qwen25-coder-7b-verilog-direct-baseline-lora	v9	Qwen2.5-Coder-7B-Instruct	Stable direct LoRA baseline trained from auto-distilled RTL data and passing anchors.
qwen25-coder-7b-verilog-delta-distilled-lora	v30b	Qwen2.5-Coder-7B-Instruct + v9	Delta-distilled direct adapter emphasizing cases where v9 failed and a verifier/selector found a passing candidate.
qwen35-9b-verilog-migration-general-lora	v32	Qwen3.5-9B	First Qwen3.5 migration adapter trained from Qwen2.5-derived verified data.
qwen35-9b-verilog-thinking-reinforced-lora	v33	Qwen3.5-9B + v32	Thinking/reasoning-preserving direct adapter with failure-analysis reinforcement and passing anchors.
qwen35-9b-verilog-structured-repair-lora	v34	Qwen3.5-9B + v33	Structured/manual repair adapter focused on deterministic final-code formatting and repair behavior.

v36 in the paper is not a separate adapter. It is a verification-guided workflow that composes the v33 and v34 adapters with compiler/simulator feedback under a bounded call budget.

Loading example

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base_model = "Qwen/Qwen2.5-Coder-7B-Instruct"  # for v9/v30b artifacts
adapter_dir = "qwen25-coder-7b-verilog-direct-baseline-lora"

tok = AutoTokenizer.from_pretrained(adapter_dir)
base = AutoModelForCausalLM.from_pretrained(base_model, device_map="auto")
model = PeftModel.from_pretrained(base, adapter_dir)

Use Qwen/Qwen3.5-9B as the base family for the Qwen3.5 adapters (v32, v33, v34) with the compatibility loader used in the project scripts.

Reporting note

These adapters are research artifacts. Several training rows are benchmark-targeted or verifier-derived. Do not report them as clean zero-shot leaderboard submissions on the same benchmark. Direct adapter results and verification-guided workflow results should be reported separately.

Internal diagnostic benchmark

This repository also includes internal-rtl-diagnostic-30/, the Internal RTL Diagnostic-30 suite used for local regression testing and adapter sanity checks. It is an internal diagnostic benchmark, not a public leaderboard. Results on this suite should not be compared to external adapters/models unless prompts, harness, extraction, simulator, and scoring protocol are matched.