Base model and loading
This repository contains adapter weights, not the full 27B model.
Use Qwen/Qwen3.8-27B at revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0,
then load this repository with peft.PeftModel.from_pretrained.
The training environment uses Transformers 5.22.1, PEFT 0.21.2, and PyTorch 2.8.0.
The base is loaded through Qwen3_5ForConditionalGeneration, in BF16.
Use the included chat template with enable_thinking=False. The evaluation protocol uses greedy decoding
and max_new_tokens=4096.
Training
LoRA rank 16, alpha 32, dropout 0; maximum sequence length 4,096; gradient accumulation 16;
peak learning rate 5e-5 with 5% warmup. The clean training candidate contains 22,260 rows,
of which 22,123 fit the sequence limit. Training exposure is 70% public-source-derived examples
and 30% original generated examples; this is a sampling ratio, not a row-count ratio.
Sources include QCircuitBench, Agent-Q, QASM-Eval, MQT Bench, QASMBench,
Qly verified OpenQASM circuits, and quantum-circuits-21k. PQID is not included in this run.
Source-derived examples underwent filtering and task-specific repair/checking; checker coverage is not universal.
No training dataset or optimizer state is redistributed in this upload. Upstream datasets retain their own terms;
the repository license does not relicense those datasets.
Evaluation and limitations
Checkpoint selection uses 2,479 validation examples. The final held-out test has 2,494 examples
and is not used for checkpoint selection. Neither evaluation is complete for this checkpoint.
Training loss is not an evaluation result. Hardware timing/pulse behavior is not
universally certified by the available checkers. Generated circuits must be checked before use.
training_provenance.json records the source snapshot hashes and training contract.