Qwen3-4B-GRPO-DeepMath-50-steps

A Qwen3-4B checkpoint used to provide on-policy parameter update directions for OPSFT. This is the GRPO on DeepMath variant, saved at global step 50.

Model Details

• Base model: Qwen/Qwen3-4B
• Role: Provides on-policy parameter update directions consumed by the OPSFT framework
• Training: GRPO on DeepMath, global_step 50
• Precision: bfloat16
• Architecture: Qwen3ForCausalLM

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("shufanshen/Qwen3-4B-GRPO-DeepMath-50-steps")
model = AutoModelForCausalLM.from_pretrained("shufanshen/Qwen3-4B-GRPO-DeepMath-50-steps", torch_dtype="bfloat16")

Intended Use

Research on on-policy distillation, reinforcement learning, and parameter update direction extraction. Not a consumer product; not for deployment.

License

Distributed under the Qwen License Agreement. Base model weights are subject to Qwen's license.