{ "command": "train", "experiment": "/data-vol1/sangwooA/sclm/experiments/qwen3_4b_rewind_boundary_sft_v1", "source": "/data-vol1/sangwooA/sclm/experiments/qwen3_4b_rewind_boundary_data_v1", "steps": 100, "batch_size": 8, "earlier_samples": 32, "lr": 0.001, "warmup_steps": 5, "save_every": 10, "seed": 42, "arm": "timing", "training_labels": "Existing positive/negative gain labels unchanged; no gain-tie examples added", "negative_loss_weight": 1, "preservation": { "input_embeddings_unchanged": true, "ordinary_output_rows_unchanged": true, "backbone_frozen_entire_run": true, "input_output_untied": true, "trainable_parameters": 2560, "token_id": 151669 }, "runtime": "Environment must intercept , delete four sections, refund erased tokens, resume with a native prompt and no marker, and block further rewinds after one use.", "waiting_loss_weight": 1, "loss": "mean[-log p(target) + sum_before_target -log(1-p)] + mean gain-negative -log(1-p)", "earlier_labels": "Timing targets, not claims of negative rewind value. May include unknown/tied/other positive positions when waiting for a later sampled target." }