rl_agent_config.json

{
  "encoder": "answerdotai/ModernBERT-large",
  "head_layers": 2,
  "max_len": 1024,
  "head_max_len": 256,
  "max_prefixes": 6,
  "act_costs": {
    "escalate": 0.5
  },
  "cost_wrong_act": 3.0,
  "amp_dtype": "bf16",
  "model_name": "enterprise-reflux-laya",
  "temperature": [
    3.50919508934021,
    1.0,
    1.0
  ],
  "temperature_by_options": {},
  "training": {
    "recipe": "Laya RLCD + soft CE",
    "base_model": "convaiinnovations/laya",
    "epochs": 4,
    "updates": 3696,
    "encoder_lr": 2.5e-05,
    "head_lr": 0.0001,
    "group_size": 4,
    "sigma_start": 0.4,
    "sigma_end": 0.1,
    "v1_train_repeat": 2,
    "calibration_held_out_before_training": true
  },
  "gradient_checkpointing": true,
  "max_tokens_per_batch": 4096,
  "fine_tuned": true
}