Qwen3.5-9B-eSFT GSM8K no-CoT SFT
This checkpoint is a supervised fine-tune of zhengbang0707/Qwen3.5-9B-eSFT on GSM8K. Targets contain the final answer only (no chain-of-thought). Training uses Qwen3.5's multimodal conditional-generation architecture, the native chat template with enable_thinking=False, BF16 parameters, and 4-GPU FSDP. The final model was consolidated without intermediate optimizer checkpoints.
- Source revision:
2e191dfb3e32bdaaa1f0847ec3ec91a99c76595f - Dataset:
openai/gsm8k,maintrain split; validation loss on the held-out split - Training: 3 epochs, final-only BF16 checkpoint
- Final training loss: 0.4156
- Final validation loss: 0.7487
- Bedrock-only judge evaluation: 1319 examples, 503 judged correct (38.13%)
- Judge model:
us.anthropic.claude-sonnet-4-6inus-east-2
Usage
Load with AutoModelForMultimodalLM and AutoProcessor from Transformers. Keep the Qwen3.5 native processor/chat-template path and set enable_thinking=False for no-CoT inference.