This repository contains the Qwen3-4B Chemical GRPO batch-size-32 run. The repository name uses the requested GRPO-TR suffix, but the actual training method for this checkpoint is GRPO.

The repository root contains the best validation checkpoint (global_step_100), selected by validation mean@16. checkpoints/last/ contains the final checkpoint. For this run, best and final are the same checkpoint.

Performance

Dataset	Method	Base model	Train batch size	Best val mean@16	Best checkpoint	Final val mean@16	Final checkpoint
Chemical	GRPO	Qwen3-4B	32	68.81%	100	68.81%	100

Raw result files:

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

Validation Mean@16

step	val_mean16
10	0.439880952381
20	0.505952380952
30	0.552678571429
40	0.590476190476
50	0.631845238095
60	0.648214285714
70	0.655952380952
80	0.665773809524
90	0.678869047619
100	0.688095238095

Training Hyperparameters

Hyperparameter	Value
Base model	`Qwen/Qwen3-4B`
Method	`GRPO`
Train batch size	`32`
Train max samples	`3200`
Total training steps	`100`
Save frequency	`10`
Test frequency	`10`
Rollout samples per prompt	`8`
Validation samples per prompt	`16`
Learning rate	`1e-6`
vLLM GPU memory utilization	`0.8`
Parsed train score rows	`100`

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "SeongryongJung/Qwen3-4B-Chemical-GRPO-TR"
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,
)

Source

Checkpoint: checkpoints/datasets/sciknoweval/chemistry/qwen3gen-chemistry-GRPO-Qwen-Qwen3-4B-mbs8-train32-rollout8-lr1e-6-vllm0.8
W&B run: run-20260702_050732-0ozhfamy
Full queue log: artifacts/queue.log