This repository contains the last checkpoint (global_step_100) for qwen3gen-material-SRPO-Qwen-Qwen3-4B-mbs32-ema0.05-dwtrue-train64-rollout8-lr5e-6-vllm0.8, converted to Hugging Face Transformers format.

Evaluation

The reported headline score is the best validation mean@16 observed during training. It is not necessarily the score of the uploaded last checkpoint.

Dataset	Method	Model	Uploaded checkpoint	Best val mean@16	Best step	Final val mean@16
material	SRPO	Qwen3-4B	global_step_100	79.72%	90	74.20%

Raw result files:

results/validation_mean16.csv
results/training_scores.csv
artifacts/config.yaml
artifacts/wandb-summary.json

Training Setup

Base model: Qwen/Qwen3-4B
Dataset: material
Method: SRPO
EMA teacher update rate: 0.05
Uploaded weights: last checkpoint, global_step_100
Validation metric used for headline score: val-aux/*/mean@16
Validation sampling: n=16

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "SeongryongJung/qwen3-4b-material-srpo-ema005"
tokenizer = AutoTokenizer.from_pretrained(repo_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    repo_id,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)

Notes

Most best-validation intermediate checkpoints were not retained as full actor checkpoints because training kept only the latest actor checkpoint. Therefore, this repository publishes the last checkpoint and records the best validation score separately.