seccodeplt-qwen2.5-coder-7b-fixed-mixed-grpo-alpha-0.5-pi-theta-real-detector-reward-v3
Fixed Mixed GRPO trainable pi-theta (alpha=0.5) 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 | 0.585983 |
| Output format pass | 99.39% |
| Syntax pass | 98.17% |
| Capability pass | 38.41% |
| Safety pass | 62.20% |
| Detector clean | 60.37% |
| Detector score | 0.782317 |
| Joint pass | 31.10% |
Important: this is pi-theta
This repository stores the trainable pi-theta checkpoint, not a statically
merged policy. Reproduce the evaluated policy by mixing this model's logits
with the frozen anchor xw1234gan/seccodeplt-qwen2.5-coder-7b-diff-sft-v2 using alpha=0.5:
mixed_logits = 0.5 * pi_theta_logits + 0.5 * anchor_logits
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.