router-expert-grade_science

LoRA expert for the grade_science target of the route-then-admit pool, trained on meirdick/router-experts-data data/experts/grade_science.jsonl.

• base: Qwen/Qwen3-4B-Instruct-2507
• rank 16, alpha 32, dropout 0.05, modules q_proj, k_proj, v_proj, o_proj
• lr 0.0002, epochs 2, max_len 1024, token budget 8192 per batch
• one example per item per recipe (direct,cot_short), under the serving system prompt; 40% of the mc items re-lettered to 5 to 10 options
• loss on the assistant turn only; held-out 5% for the val loss

field	value
target	grade_science
items	1500
recipes	direct,cot_short
kinds	{'mc': 1500}
mc_padded	534
examples	3000
encoded	3000
dropped_too_long	0
train_rows	2850
val_rows	150
train_batches	51
train_tokens	403963
supervised_tokens	64914
steps	102
train_loss	0.8913410418100801
val_loss_before	4.000902233691571
val_loss	0.8184471968267087
seconds	185.67
trainable_params	11796480