matformer_config.json

{
  "model_class": "Autoregressive_Model",
  "model_config": {
    "name": "BabyLM",
    "hidden_size": 768,
    "ffn_factor": 3.0,
    "num_hidden_layers": 12,
    "num_attention_heads": 12,
    "vocab_size": 32777,
    "bos_token_id": 1,
    "eos_token_id": 2,
    "pad_token_id": 0,
    "tie_word_embeddings": false,
    "rms_norm_eps": 1e-06,
    "attention_type": [],
    "max_position_embeddings": 1024,
    "block_size_for_attention": 128,
    "rope_theta": 10000.0,
    "compile_flexattn": false,
    "bias": false,
    "training_objective": "autoregressive",
    "is_causal": true,
    "default_layer": {
      "attn_impl": "flash",
      "sliding_window_size": null,
      "positional_encoding": "alibi",
      "normalization": "rmsnorm",
      "normalization_position": "post",
      "ffn_activation": "swiglu",
      "hooks": {}
    },
    "custom_layers": {}
  },
  "training": {
    "optimizer": "muon",
    "lr_scheduling": true,
    "lr": 0.0008,
    "final_lr": 8e-05,
    "hold_steps": 0.013,
    "weight_decay": 0.01,
    "scheduler": "custom",
    "gradient_clip_val": 1.0,
    "warmup_steps": 0.029,
    "max_epochs": 10,
    "accumulate_grad_batches": 5,
    "seed": 27,
    "save_every_n_steps": 8000,
    "checkpoint_name": "Baby_modello_di_prova_28L"
  },
  "tokenizer": {
    "type": "huggingface",
    "pretrained_name": "sapienzanlp/Minerva-350M-base-v1.0",
    "varlen_strategy": "unpadding"
  },
  "data": {
    "data_root": "/home/matteo/BabyLM/mdat_baby_finale/",
    "batch_size": 36,
    "num_workers": 1,
    "mdat_strategy": "MinervaSpecialTokens",
    "mdat_view": null
  },
  "save_dir": "./checkpoints_baby",
  "wandb_project": "BabyLM",
  "wandb_run_name": "Baby-Test-Model-Piccolo"
}