config.json

{
  "model_type": "vit-captioner-bias-decoder",
  "feature_extractor": "google/vit-base-patch16-224-in21k",
  "vocab_size": 75460,
  "seq_len": 32,
  "feature_dim": 768,
  "training_epochs": 3,
  "dataset": "ROCO-radiology (train + val + test)",
  "trainable": "Decoder + ViT biases only",
  "description": "ROCO radiology captioner trained for 3 epochs on full dataset using cached ViT features."
}