mssfj/qwen25-0.5b-finemath-4plus-openmath-sft

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Oct 6, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The mssfj/qwen25-0.5b-finemath-4plus-openmath-sft is a 0.5 billion parameter Qwen2.5-based language model, fine-tuned by mssfj for mathematical instruction-following. It was trained on a specialized dataset derived from OpenMathInstruct-2, focusing on MATH and GSM8K problems, and uses a 32768 token context length. This model is specifically designed as a mathematical baseline for low-bit math reasoning comparisons and is intended for controlled mathematical experiments rather than general-purpose assistant use.

Loading preview...

Overview

This model, mssfj/qwen25-0.5b-finemath-4plus-openmath-sft, is a 0.5 billion parameter supervised fine-tune of the Qwen2.5 architecture, developed by mssfj. It builds upon a base model pretrained on FineWeb-Edu and further on FineMath 4+. The primary purpose of this model is to serve as a mathematical instruction-following baseline for research in low-bit mathematical reasoning.

Key Capabilities

  • Mathematical Instruction Following: Fine-tuned specifically to solve math problems step-by-step, expecting a final answer in a \boxed{ANSWER} format.
  • Specialized Training: Utilizes a curated dataset, mssfj/qwen25-openmathinstruct2-sft-50k, which combines 70% MATH and 30% GSM8K problems, with careful filtering to minimize overlap with test sets.
  • Qwen ChatML Format: Processes prompts and generates responses using the Qwen ChatML format, with loss calculated only on assistant response tokens.

Benchmark Results

Evaluated on October 6, 2026, the model achieved:

  • GSM8K: 12.21% accuracy (161 correct out of 1,319 questions).
  • MATH500: 4.80% accuracy (24 correct out of 500 questions), including an evaluator's retry mechanism.

Good For

  • Mathematical Research: Ideal for controlled experiments and comparisons in mathematical reasoning, especially within the context of low-bit model performance.
  • Understanding SFT Baselines: Provides a clear baseline for how a small, specialized model performs on common mathematical benchmarks after supervised fine-tuning.

Limitations

This model is intended for controlled mathematical experiments and not for general-purpose assistant use. The presence of a boxed answer is a trained convention and does not guarantee mathematical correctness.

Finetunes
2 models