sft_freeze/train_args.json

{
  "model_name": "Qwen/Qwen3.5-0.8B",
  "init_adapter": "checkpoints/qwen35_history_actions_eosK_aw2_accum3_from_step210/current",
  "out": "checkpoints/qwen35_history_actions_eosK_aw2_accum4_from_step140",
  "steps": 1000,
  "quota": "tp=5,k=2,scroll=2,wait=2,system=1",
  "lr": 0.0001,
  "grad_accum": 4,
  "optim": "paged_adamw_8bit",
  "action_weight": 2.0,
  "margin_weight": 0.25,
  "margin_target": 8.0,
  "history_len": 4,
  "min_history": 4,
  "hist_width": 232,
  "hist_gutter": 16,
  "curr_width": 464,
  "curr_gutter": 32,
  "max_seq_len": 2600,
  "max_raw_tree_chars": 4000,
  "max_pix_rows": 12000,
  "max_mm_tokens": 2600,
  "save_every": 5,
  "log_every": 1,
  "log_shapes": true,
  "skip_oom": true,
  "shards": [
    0,
    1,
    2,
    3,
    4,
    5,
    6,
    7,
    8,
    9,
    10,
    11,
    12,
    13,
    14,
    15,
    16,
    17
  ],
  "seed": 3396423176,
  "fresh_seed": true
}