training_config.json

{
  "output_dir": "models/braille256_scaled/braille256_scaled_64M",
  "learning_rate": 0.0003,
  "weight_decay": 0.01,
  "warmup_steps": 1000,
  "max_steps": 10000,
  "per_device_train_batch_size": 16,
  "per_device_eval_batch_size": 16,
  "gradient_accumulation_steps": 2,
  "max_seq_length": 512,
  "logging_steps": 100,
  "eval_steps": 500,
  "save_steps": 1000,
  "track_emergent_patterns": true,
  "pattern_analysis_steps": 1000,
  "fp16": true,
  "dataloader_num_workers": 4
}