xw1234gan/seccodeplt-qwen2.5-coder-7b-diff-sft-v2
The xw1234gan/seccodeplt-qwen2.5-coder-7b-diff-sft-v2 is a 7.6 billion parameter Qwen2.5-Coder-7B-Instruct model, fine-tuned by xw1234gan using token-diff supervised fine-tuning for the SecCodePLT+ compliance experiment. This version corrects causal-label alignment and utilizes the ReaL safety-unit-test reward with DAPO-style token loss and dynamic sampling. It is specifically optimized for generating secure code, demonstrating a 64.02% safety pass rate and a 31.71% joint pass rate on official test examples. This model is ideal for developers focused on generating compliant and secure Python code within resource-bounded environments.
Loading preview...
Model Overview
This model, seccodeplt-qwen2.5-coder-7b-diff-sft-v2, is a specialized fine-tuned version of the Qwen/Qwen2.5-Coder-7B-Instruct base model, developed by xw1234gan. It leverages token-diff supervised fine-tuning (SFT) specifically for the SecCodePLT+ compliance experiment. The v2 iteration incorporates crucial improvements, including corrected causal-label alignment and the integration of the official ReaL safety-unit-test reward, utilizing DAPO-style token loss and dynamic sampling during training.
Key Capabilities & Performance
This model is designed to enhance code security and compliance. Its evaluation on 164 official test examples, using greedy decoding, yielded the following metrics:
- Mean reward: 0.514540
- Output format pass: 99.39%
- Syntax pass: 98.78%
- Capability pass: 39.02%
- Safety pass: 64.02%
- Joint pass: 31.71%
Intended Use & Limitations
This model is particularly suited for tasks requiring the generation of secure and compliant Python code, especially within contexts where adherence to specific safety and format standards is critical. It was trained with seed 42 on a 655-example training split. It's important to note that this is a single-seed research checkpoint, evaluated using a benchmark's resource-bounded Python verifier. Therefore, while it demonstrates strong performance in its specialized domain, it does not offer a general guarantee of secure code for all possible scenarios.