PreciseCoder-27B

Qwen3.6-27B post-trained with reinforcement learning for precise debugging: given buggy Python code, produce a minimal, correct fix (PreciseCoder project). Best-validation checkpoint, global step 328 (2 epochs).

Recipe. GDPO from the base model (no SFT warm-start): two reward signals, unit-test pass (weight 2.0) and unit-gated F1 (weight 1.0), each normalized within the rollout group before being combined into the advantage. F1 is the edit-block precision/recall F1 against the ground-truth diff (PDB evaluator), counted only when the unit tests pass. Trained on the PDB train split (buggy inputs only) with verl on 1 node of 8×H200 (n=6 rollouts, temperature 1.0, lr 1e-6, batch 32, max response 6144 tokens, KL loss 0.005, rollout TP=2, token-level rollout importance sampling). Minimal-edit prompt template.

Eval (PDB benchmark, BigCodeBench + LiveCodeBench debugging tasks):

eval	result
pdb_test (buggy inputs, 4k think + 2k answer)	unit 0.793 · precision 0.792 · recall 0.867 · F1 0.810 · edit lines 2.6
gt_holdout (1512 clean inputs, 16k think + 4k answer)	no-op 45.8% · damage 11.1% · unit pass 89.0%
PDB-Wild (228 SWE-smith repo-level bugs)	unit 0.522 · precision 0.551 · recall 0.629 · F1 0.566 · edit lines 9.2

no-op = share of clean (bug-free) inputs returned unchanged; damage = share broken by an edit.

Load with transformers / vLLM (>=0.17) as a standard Qwen3.6 model (thinking enabled by default). Source: verl experiment qwen36-27b-GDPO_u2f1_f1ur_h200 (Delta, Oct 2026). The previous release (GRPO, reward (unit + 0.5·F1)/1.5) is at revision 82cf7e4.