name: think_baseline
base_model: Qwen3-4B-Base
method: full SFT (ZeRO-3)
dataset: think_all
template: qwen3
epochs: 2
seed: 42
cutoff_len: 24576
learning_rate: 2e-5
train_loss: 0.6698
train_steps: 838
finished_at: 2026-06-08 00:22 CST