Results (held-out test set, 2,494 prompts)
Greedy decoding, max_new_tokens=4096. A prompt counts as correct only when the generated circuit passes a semantic
check (operator, state or distribution equivalence), not a text match.
Table with columns: Model, Correct, Rate| Model | Correct | Rate |
|---|
| clean-v12 step 1000 (starting point) | 2,341 | 93.9% |
| clean-v13 step 600 (this release) | 2,411 | 96.7% |
On the same prompts, 93 more are correct than for the starting point.
Where it improved
Table with columns: Task family, Prompts, step 1000, step 600| Task family | Prompts | step 1000 | step 600 |
|---|
| mqt_full_adder | 12 | 0 | 12 |
| mqt_half_adder | 5 | 0 | 4 |
| qasm_eval_complex | 31 | 2 | 6 |
| mqt_qft | 38 | 17 | 19 |
| agentq_qaoa | 281 | 262 | 272 |
| qaoa | 347 | 332 | 347 |
| teleportation | 198 | 189 | 198 |
| reset_feedback | 184 | 180 | 184 |
| ghz | 149 | 144 | 149 |
| phase_oracle | 66 | 61 | 65 |
| qasm_eval_timing | 31 | 27 | 29 |
| product_ry | 312 | 310 | 312 |
Adder widths in the test set do not appear in the training data.
Loading
The repository contains adapter weights only. Use Qwen/Qwen3.8-27B at revision
1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 (loaded with Qwen3_5ForConditionalGeneration, BF16),
then peft.PeftModel.from_pretrained. Use the included chat template with enable_thinking=False.
Environment: Transformers 5.22.1, PEFT 0.21.2, PyTorch 2.8.0.
Training
Continued from the clean-v12 step-1000 adapter: 600 steps, learning rate 2e-5, gradient accumulation 16,
LoRA rank 16 / alpha 32, maximum length 4,096, BF16.
Data: the 22,260 clean-v12 rows plus 2,333 added rows (24,593 in total). The added rows are verified
arithmetic and QFT circuits (adders, comparators, multipliers, QFT, phase estimation) and circuits generated
from MQT Bench and the Qiskit circuit library. Every added circuit was checked against its specification and
screened for overlap with the validation and test sets. Training data is not redistributed here;
upstream datasets keep their own terms.
The earlier clean-v12 adapters are in LeonaQuantum/leona-qasm-1.0-beta.
eval_results.json holds the numbers above; training_provenance.json holds the training contract and file hashes.