YumengWithG/Qwen3.5-9B-meSFT-gsm8k-noCoT-SFT

VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 13, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

YumengWithG/Qwen3.5-9B-meSFT-gsm8k-noCoT-SFT is a 9 billion parameter Qwen3.5-based model, fine-tuned specifically for mathematical reasoning on the GSM8K dataset. This model is optimized to provide direct answers to math problems without generating chain-of-thought explanations. It leverages Qwen3.5's multimodal conditional-generation architecture, making it suitable for focused numerical problem-solving tasks.

Loading preview...

Overview

This model, YumengWithG/Qwen3.5-9B-meSFT-gsm8k-noCoT-SFT, is a specialized supervised fine-tune of zhengbang0707/Qwen3.5-9B-meSFT. Its primary focus is on mathematical reasoning, specifically trained on the GSM8K dataset to produce final answers without intermediate chain-of-thought (CoT) steps.

Key Capabilities

  • Direct Mathematical Problem Solving: Optimized to provide concise, final answers for arithmetic and word problems, bypassing detailed reasoning steps.
  • Qwen3.5 Architecture: Built upon the Qwen3.5 multimodal conditional-generation architecture, utilizing its native chat template.
  • Efficient Inference: Configured for no-CoT inference by setting enable_thinking=False in the processor.

Training Details

  • Dataset: Fine-tuned using the openai/gsm8k dataset's main training split.
  • Parameters: Trained with BF16 parameters over 3 epochs using 4-GPU FSDP.
  • Performance: Achieved a final training loss of 0.4120 and a validation loss of 0.7454. Judge evaluation on 1319 examples showed 37.38% correctness using us.anthropic.claude-sonnet-4-6 as the judge model.

When to Use This Model

This model is ideal for applications requiring quick, direct numerical answers from mathematical problems, particularly when explicit step-by-step reasoning is not needed or desired. It's suitable for integration into systems where only the final numerical output is consumed.