seccodeplt-qwen2.5-coder-3b-grpo-no-kl-real-detector-reward-v3
GRPO without KL regularization for the SecCodePLT+ compliance experiment using
Qwen/Qwen2.5-Coder-3B-Instruct. This v3 run uses ReaL's program-analysis detector reward
with DAPO-style token loss and dynamic sampling. The reward is 0.5 * capability_test_fraction + 0.5 * max(0, 1 - 0.3 * detected_vulnerabilities).
Training used seed 42 and the official 655-example training split.
Evaluation used greedy decoding on all 164 official test examples.
Evaluation
| Metric | Value |
|---|---|
| Mean reward | 0.511832 |
| Output format pass | 97.56% |
| Syntax pass | 97.56% |
| Capability pass | 25.61% |
| Safety pass | 56.10% |
| Detector clean | 54.27% |
| Detector score | 0.768293 |
| Joint pass | 17.07% |
Limitations
This is a single-seed research checkpoint evaluated with the benchmark's resource-bounded Python verifier. It is not a general guarantee of secure code.