YumengWithG/Qwen3.5-9B-eSFT-gsm8k-noCoT-SFT
YumengWithG/Qwen3.5-9B-eSFT-gsm8k-noCoT-SFT is a 9 billion parameter Qwen3.5-based language model, specifically fine-tuned for mathematical reasoning on the GSM8K dataset. This model is optimized to provide direct answers without chain-of-thought (CoT) reasoning, leveraging Qwen3.5's multimodal conditional-generation architecture. It excels at arithmetic problem-solving where only the final answer is required, making it suitable for applications needing concise mathematical outputs.
Loading preview...
Model Overview
YumengWithG/Qwen3.5-9B-eSFT-gsm8k-noCoT-SFT is a 9 billion parameter model derived from zhengbang0707/Qwen3.5-9B-eSFT. It has been specifically fine-tuned on the GSM8K dataset to generate direct answers to mathematical problems, omitting the typical chain-of-thought (CoT) reasoning steps. This model utilizes Qwen3.5's multimodal conditional-generation architecture and was trained using BF16 parameters and 4-GPU FSDP.
Key Characteristics
- Parameter Count: 9 billion parameters.
- Fine-tuning: Supervised fine-tuning on the GSM8K dataset, focusing on final answer generation.
- Architecture: Based on Qwen3.5's multimodal conditional-generation architecture.
- Inference: Designed for "no-CoT" inference, providing concise, direct answers.
- Performance: Achieved a final training loss of 0.4156 and a validation loss of 0.7487. Evaluated with a 38.13% correctness rate on 1319 examples using
us.anthropic.claude-sonnet-4-6as a judge model.
Usage Considerations
This model is best suited for tasks requiring direct, final answers to mathematical problems, particularly those similar to the GSM8K dataset. When loading with AutoModelForMultimodalLM and AutoProcessor from Transformers, ensure enable_thinking=False is set for no-CoT inference to align with its fine-tuning objective.