moofeez/qwen3-coder-30b-a3b-debugger-rl

Qwen3-Coder-30B-A3B debugger, RL adapters

RL policy adapters from llm-debugger, a model that works a failing Python test in a live pdb session and edits the fix. Both are rank-16 DoRA RFT deltas over Qwen/Qwen3-Coder-30B-A3B-Instruct, trained with GRPO.

dir	headline	servable
v90/	val 75.9% pass@1 / 93.1% pass@3 (40 tasks, 8 runs)	no
v144/	heldout 62.8% pooled (43 tasks x 3)	only with its frozen SFT layer

Before loading either

Neither is a complete model, and neither can be merged onto the raw base. Each is a delta over a frozen SFT adapter. Its adapter_config.json names the raw base, so merging it directly produces a model that never existed and any number measured that way is meaningless. The stack is base -> frozen SFT adapter -> RFT delta; evaluation/scripts/merge_rl_policy_for_eval.py does it.

v90/ cannot be served at all. Its frozen layer is gone — it existed only on the cluster the campaign ran on, was never committed, uploaded, or logged as an artifact. What survives: the training corpus (llm-debugger-sft-corpus, 312 rows), the recipe (seed 42, DoRA rank 24, 4 epochs, lr 1e-4, effective batch 8 at 16k), and the target to accept a reproduction by (best_val_loss 0.11644 at 156 steps, 432 LoRA tensors). Retraining restores the init lineage, not v90 — a delta is bound to the weights it trained against.

On the numbers

Cite v90's val figure (75.9 / 93.1), not its test figure. The campaign gated on the set that is the pristine test split under defect_split_v2.json, fired nine times, and picked v90 by stop-at-peak:

v0 64.4 | v15 65.6 | v30 63.3 | v45 70.0 | v60 62.2 | v75 72.2 |
v90 76.7 | v105 68.9 | v120 66.7

76.7 is the argmax of nine noisy draws on the set it was then scored against; the drop to 72.9 at 8 runs per arm is that resolving. Validation has zero overlap with test and did not select the checkpoint. For a matched-power SFT-to-RL delta on test, the preceding campaign is clean: 60.8/81.5 -> 69.2/85.5, both at 8 runs.

Solve rate is final_test_passed, not anything the model asserts.

Training

GRPO on a 2xH200 profile: adaptive reference-KL leash, turn-budget curriculum, behaviour-cloning anchor and replay seed on the SFT stage corpus. Rewards are outcome-dominated; debugger adoption is deliberately not an objective, because debugger-first incentives drove wrong-site fixation on 16 of 16 regression scenarios. Definitions in rl_trainer/environment/rewards/.

Transcripts for every reported number: llm-debugger-eval-transcripts.