jepetolee/Qwen3-4B-AMQ3-Math-SFT

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 21, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

jepetolee/Qwen3-4B-AMQ3-Math-SFT is a supervised fine-tuned Qwen3-4B-Base model, developed by jepetolee, specifically optimized for mathematical reasoning tasks. It was trained on approximately 292,000 short chain-of-thought math solutions distilled from Qwen3-235B-A22B. This model serves as a clean math-reasoning cold-start checkpoint, intended for research purposes only due to its training data's license restrictions.

Loading preview...

Overview

This model, jepetolee/Qwen3-4B-AMQ3-Math-SFT, is a supervised fine-tune of the Qwen/Qwen3-4B-Base architecture. Its primary focus is on mathematical reasoning, achieved by training on a dataset of approximately 292,375 short chain-of-thought (CoT) math solutions. These solutions were distilled from the larger Qwen3-235B-A22B model, making it a specialized checkpoint for math-related tasks.

Key Capabilities & Features

  • Specialized Math Reasoning: Fine-tuned exclusively on math problems with CoT solutions, aiming for accurate step-by-step reasoning.
  • Qwen3 Chat Template: Utilizes the official Qwen3 chat template, expecting a specific prompt format including <think> tags for reasoning and \boxed{} for answers.
  • Optimized Generation with Logits Processors: Includes custom vLLM logits processors to manage generation, such as:
    • Think-budget cut: Enforces a maximum token budget for the reasoning (<think>) section (e.g., 6144 tokens out of 10240 total), ensuring an answer is produced.
    • Repetition abort: Prevents runaway generations by detecting and stopping n-gram repetitions.
  • Recommended Sampling: Optimal performance is achieved with temperature <= 0.75 for clean generations, though the custom logits processors allow for higher temperatures.

Limitations

  • Math Only: Designed exclusively for English math problems; MCQ items were filtered from training data.
  • Boxed Answers: Expects and produces answers within \boxed{}.
  • Research Use Only: Due to the license of its distillation data, this model is restricted to research purposes only and cannot be used for commercial or potentially harmful applications.

Finetunes
1 models