drawais/circuitqa-adapters

CircuitQA LoRA adapters

LoRA adapters (r = 64, alpha = 128; weights stored in fp32, base model loaded in bf16) for Qwen/Qwen2.5-Coder-7B-Instruct, trained by the authors of CircuitQA: A Grounded Architecture for Question Answering over Digital RTL Designs with Multi-Backend Query Routing. Data, code and evaluation outputs: drawais/circuitqa. Only the adapter weights and configurations are provided; the base model and its tokenizer are loaded from their own repository.

Folder	Role
s3_seed42, s3_seed123, s3_seed456	Response Decoder trained on the template questions of the repository-disjoint split (three seeds)
router_s3	Query Routing Module trained on the same template questions
s2_seed42, s2_seed123, s2_seed456	Response Decoder trained on the template questions of the design-disjoint split (three seeds)

All adapters are trained on template questions and answers computed from the outputs of the analysis backends, from an earlier build of the template corpus of drawais/circuitqa (the released corpus/ recomputes every answer with the released backends).

Requires peft>=0.21 (the configs were written by PEFT 0.21.2).

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base = "Qwen/Qwen2.5-Coder-7B-Instruct"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(model, "drawais/circuitqa-adapters", subfolder="s3_seed42")

Licence

The adapter weights and configurations in this repository are licensed under CC BY-NC 4.0 (non-commercial use). They are LoRA weights trained by the authors on top of Qwen2.5-Coder-7B-Instruct; the base model is not included and remains under its own licence, the Apache License 2.0 (copy in LICENSE-Qwen2.5-Coder-7B-Instruct). Use of the base model is subject to that licence.

Training data. The training data refer to designs from the upstream repositories listed in the dataset's LICENSE.md; the adapters contain no design source code. Training data derived from NVDLA designs contain data derived from NVDLA, Copyright (c) 2016-2017 NVIDIA Corporation, licensed under the NVIDIA Open NVDLA License and Agreement v1.0 (copy in LICENSES/), modified to contain only analysis results computed by the authors. Training data derived from files of pConst/basic_verilog written by Konstantin Pavlov refer to designs licensed under CC BY-SA 4.0, reduced to analysis results computed by the authors. Upstream projects are named only to identify the sources of the designs, without implying endorsement.

Citation

If you use the CircuitQA dataset, code or adapters, please cite:

@misc{awais2026circuitqa,
  title        = {{CircuitQA}: A Grounded Architecture for Question Answering over Digital {RTL} Designs with Multi-Backend Query Routing},
  author       = {Awais, Muhammad and Postolache, Octavian Adrian and Oliveira, Sancho Moura},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/drawais/circuitqa}},
  note         = {Dataset and code}
}

@misc{awais2026circuitqaadapters,
  title        = {{CircuitQA} Adapters: {LoRA} Adapters for {Qwen2.5-Coder-7B-Instruct}},
  author       = {Awais, Muhammad and Postolache, Octavian Adrian and Oliveira, Sancho Moura},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/drawais/circuitqa-adapters}},
  note         = {Trained model adapters}
}