config.json

{
  "base_model": "microsoft/deberta-v3-small",
  "max_length": 512,
  "num_categories": 7,
  "categories": [
    "violence",
    "hate",
    "sexual",
    "self_harm",
    "dangerous_info",
    "harassment",
    "illegal_activity"
  ],
  "pruning": {
    "layers_to_keep": [
      0,
      1,
      4,
      5
    ],
    "layers_to_drop": [
      2,
      3
    ]
  },
  "training": {
    "phase1": {
      "num_epochs": 3,
      "batch_size": 64,
      "gradient_accumulation_steps": 4,
      "learning_rate": 2e-05,
      "weight_decay": 0.01,
      "warmup_ratio": 0.05,
      "early_stopping_patience": 2,
      "best_model_metric": "f1_binary"
    },
    "phase2": {
      "num_epochs": 5,
      "batch_size": 64,
      "gradient_accumulation_steps": 4,
      "learning_rate": 2e-05,
      "weight_decay": 0.01,
      "warmup_ratio": 0.05,
      "confidence_low": 0.3,
      "confidence_high": 0.7,
      "best_model_metric": "f1_binary"
    },
    "recovery": {
      "num_epochs": 2,
      "batch_size": 128,
      "gradient_accumulation_steps": 2,
      "learning_rate": 2e-05,
      "weight_decay": 0.01,
      "warmup_ratio": 0.1
    },
    "eval_batch_size": 512,
    "num_workers": 12,
    "focal_loss_gamma": 2.0,
    "label_smoothing": 0.1,
    "category_loss_weight": 0.7,
    "asl_gamma_pos": 1.0,
    "asl_gamma_neg": 4.0,
    "asl_clip": 0.05,
    "rdrop_alpha": 1.0,
    "fgm_epsilon": 0.3,
    "ema_decay": 0.999,
    "multi_sample_dropout_count": 5
  },
  "filtering": {
    "min_confidence": 0.8,
    "dedup_similarity_threshold": 0.95,
    "min_tokens": 3,
    "max_tokens": 512,
    "target_safe_ratio": 0.55,
    "target_unsafe_ratio": 0.45
  },
  "splits": {
    "train": 0.85,
    "val": 0.1,
    "test": 0.05
  },
  "hard_negatives": {
    "model": "claude-sonnet-4-6",
    "total": 12000,
    "examples_per_request": 15,
    "max_workers": 8
  },
  "jigsaw": {
    "toxicity_threshold": 0.7,
    "max_samples": 20000,
    "use_soft_labels": false
  },
  "inference": {
    "binary_threshold": 0.3,
    "category_thresholds": {
      "violence": 0.5,
      "hate": 0.5,
      "sexual": 0.5,
      "self_harm": 0.5,
      "dangerous_info": 0.5,
      "harassment": 0.5,
      "illegal_activity": 0.5
    }
  }
}