---
license: apache-2.0
base_model: Qwen/Qwen3-4B
tags:
- cseg
- qwen3
pipeline_tag: text-generation
library_name: transformers
---

# CSEG-RM-SFT-GRPO-4B

GRPO-400 initialized from the CSEG-RM SFT checkpoint (`global_step_220`).

- Method: SFT-GRPO
- Size: 4B
- Initialization: CSEG-RM-SFT-4B (`global_step_220`)
- Checkpoint: `global_step_400`
- Base model: [Qwen/Qwen3-4B](https://huggingface.co/Qwen/Qwen3-4B)

Weights are bfloat16 Hugging Face safetensors. The judge is thinking-on and emits full JSON (`--prompt-output-spec full`).