veritasium/qwen2.5-3b-math-sft-clean-v2

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:3.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 27, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

The veritasium/qwen2.5-3b-math-sft-clean-v2 model is a 3.1 billion parameter Qwen2.5-3B-Instruct checkpoint, conservatively fine-tuned for concise integer-answer mathematics. It specializes in short mathematical derivations where the final answer is a single integer, typically formatted as \boxed{integer}. This model is intended for research into mathematical reasoning and aims to provide accurate, concise integer solutions.

Loading preview...

Overview

This model, veritasium/qwen2.5-3b-math-sft-clean-v2, is a 3.1 billion parameter Qwen2.5-3B-Instruct checkpoint specifically fine-tuned for concise integer-answer mathematics. It was developed to address regressions observed in an earlier SFT run, focusing on improving semantic correctness and output quality for mathematical tasks.

Key Capabilities

  • Specialized Mathematical Reasoning: Optimized for short mathematical derivations culminating in a single integer answer.
  • Concise Output: Designed to produce completions ending with \boxed{integer}.
  • Improved Accuracy: Aims to provide more semantically correct mathematical reasoning compared to its predecessor, with a focus on avoiding invalid arithmetic or logic.
  • Small Footprint: Based on a 3.1B parameter model, offering a balance between performance and computational efficiency.

Training and Evaluation

The model was trained on a carefully audited dataset of 190 examples, ensuring each completion contained exactly one integer answer box and no answer-forcing phrases. It underwent a paired n=1 evaluation against the base Qwen/Qwen2.5-3B-Instruct model on 831 questions. While the clean v2 model showed a slight improvement (+0.76pp) in accuracy on the clean usable slice (excluding calibration IDs), this difference was not statistically significant in the single-seed experiment. The training configuration used LoRA with rank 8 and alpha 16, targeting q_proj and v_proj modules, for 2 epochs.

Good For

  • Research in Mathematical LLMs: Ideal for researchers exploring fine-tuning techniques for mathematical reasoning in LLMs.
  • Integer-Answer Math Problems: Best suited for use cases requiring concise, integer-based solutions to mathematical problems.
  • Diagnostic Context: Useful for understanding the impact of data auditing and targeted SFT on mathematical performance.