PeetPedro/qwen2.5-coder-32b-heretic-swe-sft
PeetPedro/qwen2.5-coder-32b-heretic-swe-sft is a 32.8 billion parameter Qwen2.5-based model developed by PeetPedro, specifically fine-tuned for agentic software engineering (SWE) tasks and tool-calling. This model is designed for internal engineering use, featuring abliterated refusal behavior to enhance utility in controlled environments. It excels in resolving SWE-related problems, as indicated by its 1.0000 swebench_resolve score.
Loading preview...
Model Overview
PeetPedro/qwen2.5-coder-32b-heretic-swe-sft is a 32.8 billion parameter model built upon the PeetPedro/qwen2.5-coder-32b-instruct-heretic base, utilizing the ChatML (Qwen) format. This model represents the supervised fine-tuning (SFT) stage of an open abliterate → SFT → RFT → RLVR coding-model pipeline, specifically trained with Unsloth LoRA on agentic SWE and tool-calling data.
Key Characteristics
- Abliterated (Uncensored): Deliberately designed with reduced refusal behavior, meaning it will not reliably decline unsafe or disallowed requests. This characteristic makes it suitable for specific internal engineering applications where an independent safety layer is implemented.
- Software Engineering Focus: Fine-tuned for agentic software engineering tasks and tool-calling, aiming to provide robust support for development workflows.
- High SWE-Bench Resolution: Achieves a
swebench_resolvescore of 1.0000, indicating strong performance in resolving software engineering problems. - Training Provenance: Built using Heretic (abliteration), Unsloth, and TRL for SFT/RFT/RLVR stages, with the full pipeline details available in the heretic-coder-pipeline repository.
Intended Use Cases
This model is primarily intended for internal, gated engineering use. It should be deployed behind verify-before-merge systems and integrated with independent moderation or authorization layers. It is not recommended for direct user-facing applications without additional safety measures due to its abliterated nature.