training_summary.json

{
  "model": "llava-hf/llava-1.5-7b-hf",
  "transcoder_type": "Cross-Layer Transcoder (CLT)",
  "architecture": {
    "hidden_dim": 4096,
    "feature_dim": 8192,
    "num_layers": 31,
    "expansion_factor": 2.0
  },
  "training": {
    "steps_per_layer": 5000,
    "learning_rate": 0.0003,
    "batch_samples": 16,
    "sparsity_penalty": 0.01,
    "validation_interval": 200
  },
  "results": {
    "dead_features_pct": 0.0,
    "avg_sparsity_early_layers": "2-4%",
    "avg_sparsity_middle_layers": "3-7%",
    "avg_sparsity_late_layers": "7-29%",
    "avg_reconstruction_loss": "<0.1 for most layers"
  },
  "features": {
    "mlp_to_clt_mapping": "Included (co-activation based)",
    "decoder_weights": "Included (CLT \u2192 MLP reconstruction)",
    "mapping_shape": "[4096, 8192]",
    "description": "Each transcoder includes MLP\u2192CLT feature correlation mapping computed during training"
  },
  "usage": {
    "transcoder_files": "transcoder_L{0-30}.pt",
    "mapping_files": "mapping_L{0-30}.pt",
    "note": "Layer indices match LLaVA's language model layers (0-30 out of 31 total layers)"
  }
}