config.json

{
  "model_type": "controlmt",
  "architectures": [
    "ControlMTForSeq2SeqLM"
  ],
  "model_name": "ControlMT-v2.3",
  "trained_by": "Anand Kaman",
  "release_date": "2026-06-23",

  "vocab_size": 128000,
  "d_model": 512,
  "n_heads": 8,
  "d_ff": 2048,
  "dropout": 0.1,
  "encoder_layers_per_lang": 2,
  "decoder_layers_per_lang": 2,
  "shared_core_enc_layers": 6,
  "shared_core_dec_layers": 6,
  "max_position_embeddings": 320,
  "max_input_length": 256,

  "pad_token_id": 0,
  "bos_token_id": 1,
  "eos_token_id": 2,
  "unk_token_id": 3,

  "direction_tokens": {
    "kn2en": 4,
    "en2kn": 5,
    "rkn2kn": 12,
    "rkn2en": 13,
    "hi2en": 14,
    "en2hi": 15
  },
  "default_control_token_id": 7,

  "decoding_presets": {
    "default": {
      "method": "beam_search",
      "num_beams": 6,
      "length_penalty": 1.2,
      "no_repeat_ngram_size": 3,
      "anti_lm_alpha": 0.5,
      "max_length": 256,
      "description": "Production setting — matches reported FLORES benchmark numbers"
    },
    "fast": {
      "method": "beam_search",
      "num_beams": 4,
      "length_penalty": 1.2,
      "no_repeat_ngram_size": 3,
      "anti_lm_alpha": 0.0,
      "max_length": 256,
      "description": "~2x faster than default, ~0.5 BLEU lower"
    },
    "greedy": {
      "method": "greedy",
      "num_beams": 1,
      "max_length": 256,
      "description": "Fastest. ~1.5 BLEU lower than default. Use for high-throughput batch."
    },
    "high_quality": {
      "method": "beam_search",
      "num_beams": 8,
      "length_penalty": 1.2,
      "no_repeat_ngram_size": 3,
      "anti_lm_alpha": 0.7,
      "max_length": 256,
      "description": "~30% slower than default, marginal quality gain (~0.2 BLEU)"
    }
  },

  "training_config": {
    "framework": "PyTorch 2.x",
    "precision": "bf16 mixed",
    "optimizer": "AdamW",
    "weight_decay": 0.01,
    "lr_schedule": "warm-start fine-tune from v2.2 with low LR (1.5e-5 → 1e-5)",
    "warmup_steps": 500,
    "label_smoothing": 0.1,
    "grad_clip_norm": 1.0,
    "effective_batch_size": 96,
    "ema_decay": 0.999,
    "ema_start_step": 1000,
    "final_checkpoint": "final_v2.3.pt"
  },

  "tokenizer_class": "ControlMTTokenizer",
  "auto_map": {
    "AutoConfig": "configuration_controlmt.ControlMTConfig",
    "AutoModelForSeq2SeqLM": "modeling_controlmt.ControlMTForSeq2SeqLM",
    "AutoTokenizer": "tokenization_controlmt.ControlMTTokenizer"
  },

  "transformers_version": "4.40+"
}