train_args.json

{
  "encoder": "Alibaba-NLP/gte-modernbert-base",
  "data": "data/v4",
  "extra": [
    "data/distill/generic.jsonl"
  ],
  "out": "runs/compiled-late-gte-base",
  "lr": 5e-05,
  "head_lr": 0.0005,
  "epochs": 3.0,
  "batch_size": 32,
  "warmup": 0.05,
  "evidence_weight": 0.5,
  "score_sigma": 0.5,
  "hard_targets": false,
  "max_state_tokens": 512,
  "slots": 8,
  "limit": null,
  "log_every": 50,
  "seed": 0,
  "zero_shot_only": false,
  "arch": "compiled",
  "pair_budget": 256,
  "checkpoint_every": 500,
  "late_interaction": true
}