qwen3-4b-grpo-main

Qwen3-4B GRPO checkpoint bundle

This public repository contains the selected final/root GRPO checkpoints from the Qwen3 safety experiment. They were initialized from Qwen3-4B-Base, not from the non-Base Qwen3-4B instruct/chat model.

The bundle contains four independently loadable subdirectories:

• grpo_qwen3_4b_base
• grpo_qwen3_4b_base_general
• grpo_qwen3_4b_midtrain
• grpo_qwen3_4b_midtrain_general

Only the selected root/final model weights are included; intermediate checkpoint-* directories are intentionally omitted. Each subdirectory includes its model configuration and tokenizer files.

Example:

from transformers import AutoTokenizer, AutoModelForCausalLM

path = "Zzyy2000/qwen3-4b-grpo-main/grpo_qwen3_4b_midtrain"
tok = AutoTokenizer.from_pretrained(path)
model = AutoModelForCausalLM.from_pretrained(path, torch_dtype="auto", device_map="auto")

See the handoff code repository for the training and evaluation scripts: https://github.com/ZhengyueZhao/qwen3-midtrain-handoff