PEFT/LoRA adapter for KiCad schematic generation, to be applied on top of Qwen/Qwen3.6-27B (Apache-2.0). This is the adapter only (~152 MB); for the merged model see Ailiance-fr/SchGen-Qwen3.6-27B-EU.

Implements the SchGen method (Luo et al., 2026) on a sovereign EU stack by ailiance.

Config

LoRA rank 32, α 1024, dropout 0.01, targets q/k/v/o/gate/up/down on the top 16 layers (48–63), curriculum (3 phases) over microsoft/SchGen_dataset (MIT).

Usage

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.6-27B", torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(base, "Ailiance-fr/SchGen-Qwen3.6-27B-EU-lora")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3.6-27B")

See the merged model card for evaluation, limitations, and citation.

License

Apache-2.0. Derivative of Qwen/Qwen3.6-27B (Apache-2.0), trained on microsoft/SchGen_dataset (MIT). See NOTICE.