train_args.json

{
 "init_from": "checkpoints/jevlike-v2",
 "vision": "google/siglip-base-patch16-224",
 "train": "data/mm/train.jsonl",
 "val": "data/mm/val.jsonl",
 "text_train": "data/mix3/train.jsonl",
 "text_val": "data/mix3/val.jsonl",
 "text_frac": 0.35,
 "text_val_n": 2000,
 "out": "checkpoints/jevlike-mm-proto",
 "mm_epochs": 3.0,
 "max_steps": 0,
 "time_budget_min": 95.0,
 "max_tokens": 3072,
 "max_batch": 64,
 "grad_accum": 4,
 "lr_mm": 0.001,
 "lr_backbone": 2e-05,
 "lr_head": 0.0001,
 "weight_decay": 0.01,
 "warmup": 0.05,
 "align_steps": 200,
 "freeze_layers": 8,
 "grad_checkpointing": true,
 "no_sim_prior": false,
 "sim_dropout": 0.3,
 "pool": 2,
 "eval_every": 400,
 "log_every": 25,
 "num_workers": 5,
 "limit_train": 0,
 "limit_val": 0,
 "seed": 0,
 "no_calibrate": false
}