xw1234gan/seccodeplt-qwen2.5-coder-3b-diff-sft-v2
xw1234gan/seccodeplt-qwen2.5-coder-3b-diff-sft-v2 is a 3.1 billion parameter model based on Qwen2.5-Coder-3B-Instruct, specifically fine-tuned using token-diff supervised learning for the SecCodePLT+ compliance experiment. This model focuses on generating secure code by correcting causal-label alignment and utilizing a safety-unit-test reward with dynamic sampling. It is designed to improve code compliance and safety, particularly within a resource-bounded Python verification environment.
Loading preview...
Model Overview
xw1234gan/seccodeplt-qwen2.5-coder-3b-diff-sft-v2 is a 3.1 billion parameter language model derived from Qwen/Qwen2.5-Coder-3B-Instruct. It has undergone token-diff supervised fine-tuning (SFT) as part of the SecCodePLT+ compliance experiment. This v2 iteration addresses causal-label alignment issues and incorporates the official ReaL safety-unit-test reward mechanism, utilizing DAPO-style token loss and dynamic sampling during training.
Key Capabilities & Performance
This model is specifically optimized for generating code that adheres to safety and compliance standards. Evaluation on 164 official test examples using greedy decoding yielded the following results:
- Output format pass: 95.73%
- Syntax pass: 95.73%
- Safety pass: 55.49%
- Joint pass: 18.90%
Differentiators & Use Cases
What sets this model apart is its specialized fine-tuning for secure code generation and compliance. Unlike general-purpose code models, it explicitly integrates safety-unit-test rewards to guide its output towards more secure and compliant solutions. It is particularly relevant for scenarios requiring code generation that passes specific safety and format checks within a resource-bounded Python verification environment. This model is a research checkpoint, offering insights into token-diff SFT for security-focused code generation.
Limitations
It is important to note that this is a single-seed research checkpoint. Its evaluation was conducted with a benchmark's resource-bounded Python verifier, and therefore, it does not provide a general guarantee of secure code in all contexts.