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

VISIONConcurrent 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-posthoc-gsm8k-noCoT-SFT is a 9-billion parameter Qwen3.5-based language model fine-tuned specifically for mathematical reasoning on the GSM8K dataset. This model is optimized for direct answer generation without chain-of-thought (CoT) reasoning, utilizing Qwen3.5's multimodal conditional-generation architecture. It is designed for tasks requiring precise numerical or factual answers from mathematical problems, offering a focused approach to arithmetic and logic. The model was trained using BF16 parameters and a 32K context length, making it suitable for efficient, direct problem-solving.

Loading preview...

Model Overview

YumengWithG/Qwen3.5-9B-posthoc-gsm8k-noCoT-SFT is a specialized 9-billion parameter model derived from zhengbang0707/Qwen3.5-9B-posthoc. It has been supervised fine-tuned (SFT) on the openai/gsm8k dataset, focusing on mathematical problem-solving. A key characteristic of this model is its no-Chain-of-Thought (no-CoT) approach, meaning it is trained to produce final answers directly without intermediate reasoning steps.

Key Capabilities

  • Direct Mathematical Problem Solving: Optimized for generating final answers to arithmetic and logic problems from the GSM8K dataset.
  • Qwen3.5 Architecture: Leverages the multimodal conditional-generation architecture of Qwen3.5.
  • Efficient Inference: Designed for direct answer generation, potentially leading to faster inference for specific mathematical tasks compared to CoT models.
  • BF16 Precision: Trained with BF16 parameters for efficient computation.

Training Details

The model underwent 3 epochs of training on the GSM8K dataset's main train split, achieving a final training loss of 0.4118 and a validation loss of 0.7544. Evaluation using a Bedrock-only judge model (us.anthropic.claude-sonnet-4-6) on 1319 examples showed 38.13% correctness.

Good For

  • Applications requiring direct, concise answers to mathematical word problems.
  • Scenarios where computational efficiency is prioritized over detailed reasoning explanations.
  • Integration into systems that can parse and utilize final numerical or factual outputs from mathematical tasks.

Finetunes
1 models