William2390401/aime-gen-qwen3-4b-v3

TEXT GENERATIONConcurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 13, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

William2390401/aime-gen-qwen3-4b-v3 is a 4 billion parameter Qwen3-based instruction-tuned model specifically designed for generating novel, difficulty-calibrated AIME-style math problems. Developed by William2390401, this model excels at problem posing from a bare one-line prompt, producing problems with specified difficulty and topic. It achieves significantly improved format adherence, diversity, and calibration compared to its base model, making it ideal for educational and research applications in mathematical problem generation.

Loading preview...

Overview

William2390401/aime-gen-qwen3-4b-v3 is a 4-billion parameter Qwen3-based model, fine-tuned to generate AIME-style math problems. It's a 16-bit merge of aime-gen-qwen3-4b-lora-v3 into Qwen/Qwen3-4B-Instruct-2507. The model's core thesis is that problem-posing in LLMs is a diversity deficit, not a reasoning deficit, addressed by fine-tuning on a curated dataset.

Key Capabilities

  • AIME-style Problem Generation: Generates novel math problems in the style of the American Invitational Mathematics Examination (AIME).
  • Difficulty Calibration: Can generate problems calibrated to specific difficulty bands (e.g., problems 11-15).
  • Topic Specification: Allows users to specify the topic for problem generation (e.g., number theory).
  • Bare Prompt Input: Operates effectively with a single-line prompt, requiring no complex system prompts or few-shot examples.
  • Structured Output: Problems are output in a consistent format: <problem>…</problem><solution>…</solution><answer>N</answer>.

Performance Highlights

Compared to a fully-prompted base model, this fine-tuned version shows significant improvements:

  • Format Adherence: Achieves 63.9% adherence, a +35.0% increase.
  • Reduced Self-Duplication: Self-duplication drops to 18.3%, a −52.3% improvement.
  • Band Accuracy: Improves to 64.3%, a +4.3% increase.

Limitations

  • Validity: The model's primary limitation is its low problem validity (12% solver consensus). It is a strong problem stylist but a weak verifier, meaning it cannot reliably solve the problems it poses. This is considered an inherent limitation for a 4B model and not fixable by data.

Good for

  • Research: Ideal for studies on LLM capabilities in problem posing and educational AI.
  • Education: Useful for generating practice problems for advanced math competitions like AIME.
  • Content Creation: Developers needing to generate diverse and calibrated mathematical problems.

Finetunes
1 models