generated_by: scripts/export_upload_biology_rlsd32_tr_repo.py
note: W&B files/config.yaml was not present locally for this run; this is an export summary.
experiment: qwen3gen-biology-RLSD_TR-Qwen-Qwen3-4B-mbs8-decay0-tr0.1-train32-rollout8-lr1e-6-vllm0.8
wandb_run: run-20260702_113907-xauqhwk1
base_model: Qwen/Qwen3-4B
dataset: Biology / SciKnowEval biology
method: RLSD_TR
train_batch_size: 32
train_max_samples: 3200
total_training_steps: 100
save_freq: 10
test_freq: 10
rollout_n: 8
validation_n: 16
learning_rate: 1e-6
teacher_regularization: trust-region
teacher_update_rate: 0.1
token_reweight_lambda: 0.5
token_reweight_eps_w: 0.2
token_reweight_decay_steps: 0
vllm_gpu_memory_utilization: 0.8
best_step: 90
best_val_mean16: 0.485000000000
final_step: 100
final_val_mean16: 0.428750000000