qwen3-4b-sft-main

Qwen3-4B SFT checkpoint bundle

This public repository contains the selected balanced SFT checkpoints from the Qwen3 safety mid-training experiment. The checkpoints were initialized from Qwen3-4B-Base, not from the non-Base Qwen3-4B instruct/chat model.

Each top-level directory is an independently loadable checkpoint with its own config.json, tokenizer files, and model weights. The directory names distinguish the training data family and the standard/reasoning evaluation-balanced variants. The bundle intentionally contains selected final SFT checkpoints rather than every intermediate training checkpoint.

Example:

from transformers import AutoTokenizer, AutoModelForCausalLM

path = "Zzyy2000/qwen3-4b-sft-main"
# Select a subdirectory, for example:
# path = "Zzyy2000/qwen3-4b-sft-main/v2_doc_plus_reasoning_balanced"
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