visproj/proofkit-distilled-qwen0.5b
The visproj/proofkit-distilled-qwen0.5b is a 0.5 billion parameter Qwen2.5-0.5B-Instruct model, distilled from a 20B teacher model. Developed by visproj, it is specifically fine-tuned for generating work samples within the ProofKit application. This model demonstrates performance comparable to its 20B teacher on ProofKit's specific prompts, outperforming untuned base models and stale controls in evaluations.
Loading preview...
ProofKit Qwen 0.5B Distilled Model
This model, visproj/proofkit-distilled-qwen0.5b, is a 0.5 billion parameter variant of the Qwen2.5-0.5B-Instruct architecture. It was developed by visproj through a sequence-level data distillation process, where it was fine-tuned using completions from a larger gpt-oss-20b teacher model (visproj/proofkit-gpt-oss-20b-lora) over ProofKit's specific prompts.
Key Capabilities and Performance
- Distilled Performance: Achieves evaluation scores comparable to its 20B teacher model on held-out ProofKit prompts, with an average score of 76.6 across a 3-judge panel (Claude Opus 4.7, GPT-5.5, Qwen-3B).
- Efficiency: As a 0.5B parameter model, it offers a highly efficient solution for its specialized task, outperforming larger untuned base models and older controls.
- Purpose-Built: Designed as a core component for the ProofKit application, a work-sample generator for job seekers.
Limitations and Usage
- Prompt Format Dependency: This model is prompt-format-frozen; it is specifically trained on the exact prompt shapes used by ProofKit and will not perform optimally with reworded or free-form prompts.
- Specialized Use: It is a purpose-built component for the ProofKit app, not intended as a general chat model.
This model represents a significant improvement over earlier versions, incorporating a fix for synthetic-data leakage through "faithfulness anchors" and "seeded per-example variation" to ensure more reliable and context-aware outputs.