Qwen3.5-27B GRPO — checkpoint 440
Merged Hugging Face checkpoint from global step 440 of Qwen3.5-27B GRPO training.
- Architecture:
Qwen3_5ForConditionalGeneration - Weight precision: BF16 (
bfloat16), cast from the FP32-master merged checkpoint - Format: full model weights, split across 23 safetensors shards (not a LoRA adapter)
- Includes the checkpoint's original configuration, tokenizer, processor configuration, and chat template
The checkpoint is uploaded without quantization or changes to its weights or tokenizer.