name: think_s1_ep23
base_model: Qwen3-4B-Base (continued from checkpoint-302)
method: full SFT (ZeRO-3)
dataset: ocr_think_50k
template: qwen3
epochs: 2
seed: 42
cutoff_len: 24576
packing: true
neat_packing: false
per_device_train_batch_size: 1
gradient_accumulation_steps: 16
effective_batch_size: 64
learning_rate: 5e-5
train_loss: 0.5416
train_steps: 604
finished_at: 2026-06-10 05:23 CST