Qwen3-4B-Physics-RLSD

Qwen3-4B Physics RLSD

This repository contains Physics fine-tuned Qwen3-4B checkpoints from the local SciKnowEval-style generalization setup.

• Root checkpoint: final global_step_100 merged to Hugging Face safetensors.
• best_avg16/: checkpoint with the highest validation avg@16 during training, merged to Hugging Face safetensors.

Checkpoints

Training Run

qwen3gen-physics-RLSD-Qwen-Qwen3-4B-mbs8-decay0-ema0.05-train256-rollout8-lr1e-6-vllm0.8

Base Model

• Base model: Qwen/Qwen3-4B
• Fine-tuning type: full-parameter FSDP RL training
• Dataset: datasets/sciknoweval/physics
• Train split: 720 examples
• Validation split: 80 examples

Method

• Method: RLSD
• Config: rlsd
• Policy loss mode: rlsd
• Reward: local SciKnowEval multiple-choice reward checker
• Rollout correction: token-level importance sampling, threshold 2.0

Hyperparameters

Metrics

Loading

Root final checkpoint:

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("SeongryongJung/Qwen3-4B-Physics-RLSD")
tokenizer = AutoTokenizer.from_pretrained("SeongryongJung/Qwen3-4B-Physics-RLSD")

Best avg@16 checkpoint:

model = AutoModelForCausalLM.from_pretrained("SeongryongJung/Qwen3-4B-Physics-RLSD", subfolder="best_avg16")
tokenizer = AutoTokenizer.from_pretrained("SeongryongJung/Qwen3-4B-Physics-RLSD", subfolder="best_avg16")

Intended Use

This model is intended for research on RL fine-tuning and self-distillation behavior on science/generalization tasks. It has not been broadly safety evaluated for production use.

Limitations

The reported scores are training-time and validation-time metrics from the local experimental setup. They should not be interpreted as broad benchmark results without independent evaluation.