visproj/proofkit-qwen0.5b-7k

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jun 9, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The visproj/proofkit-qwen0.5b-7k model is a 0.5 billion parameter Qwen2.5-0.5B-Instruct variant, fine-tuned by visproj directly on approximately 7,000 synthetic examples from the ProofKit SFT dataset. This LoRA-trained and merged model specializes in ProofKit's specific task contracts, including section drafting, co-author drafting, revision actions, and strict-JSON generation for scenarios, recommendations, readiness, and portfolios. It is optimized for generating structured outputs within the ProofKit application's defined prompt formats, rather than general chat.

Loading preview...

ProofKit Qwen 0.5B (Direct SFT)

This model, visproj/proofkit-qwen0.5b-7k, is a 0.5 billion parameter version of Qwen2.5-0.5B-Instruct, specifically fine-tuned by visproj. It was trained directly on the ProofKit SFT dataset, comprising approximately 7,000 synthetic examples, using LoRA and then merging the weights.

Key Capabilities & Specialization

This model is purpose-built for the ProofKit application, a work-sample generator for job seekers. Its training focuses on specific "task contracts" within ProofKit, enabling it to:

  • Generate section drafts and co-author drafts from user input.
  • Perform revision actions on text.
  • Produce strict-JSON outputs for scenarios, recommendations, readiness assessments, and portfolio generation.

Performance & Evaluation

Despite its small size, evaluation against 15 held-out prompts by a three-judge panel (Claude Opus 4.7, GPT-5.5, and a local Qwen-3B) shows that this 0.5B model significantly outperforms untuned base models and a stale control. It achieves an average score of 75.0, demonstrating its effectiveness within its specialized domain.

Limitations & Usage

Due to its 0.5B capacity, the model reliably handles trained-style specifics but may occasionally garble novel tokens. It is not a general chat model and is highly sensitive to prompt formatting, requiring adherence to ProofKit's prompt_formats.py shapes for optimal performance. Prompting outside these specific formats will lead to off-distribution behavior.

Finetunes
2 models