Modelpublic

xw1234gan/seccodeplt-qwen2.5-coder-7b-grpo-no-kl-real-detector-reward-v3

sourceHugging Faceupdated 1mo agoView on Hugging Face

0likes115downloads

Model Card

seccodeplt-qwen2.5-coder-7b-grpo-no-kl-real-detector-reward-v3

GRPO without KL regularization for the SecCodePLT+ compliance experiment using Qwen/Qwen2.5-Coder-7B-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	0.600820
Output format pass	99.39%
Syntax pass	98.78%
Capability pass	39.02%
Safety pass	63.41%
Detector clean	61.59%
Detector score	0.798171
Joint pass	31.71%

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.