dataset_metadata.json

{
  "repository_info": {
    "name": "ericjm/narrow-data",
    "description": "Experimental model checkpoints from 'On the creation of narrow AI' paper",
    "version": "1.0",
    "paper_title": "On the creation of narrow AI: hierarchy and nonlocality of neural network skills",
    "authors": ["Eric Michaud", "Asher Parker-Sartori", "Max Tegmark"],
    "upload_date": "2024-06-22",
    "upload_method": "HuggingFace CLI"
  },
  "experiments": {
    "trainscratch01": {
      "description": "LLMs trained from scratch on GitHub code",
      "purpose": "Scaling analysis for paper Figures 6 & 12",
      "dataset": "codeparrot/github-code (Python subset)",
      "training_steps": 100000,
      "learning_rate": "5e-4",
      "sequence_length": 1024,
      "hardware": "NVIDIA A100 80GB"
    }
  },
  "models_uploaded": {
    "trainscratch01/d256_l4_h4": {
      "parameters": "23M",
      "hidden_size": 256,
      "num_layers": 4,
      "num_heads": 4,
      "intermediate_size": 1024,
      "model_size_gb": 0.15,
      "purpose": "Smallest model for scaling baseline"
    },
    "trainscratch01/d768_l12_h12": {
      "parameters": "338M", 
      "hidden_size": 768,
      "num_layers": 12,
      "num_heads": 12,
      "intermediate_size": 3072,
      "model_size_gb": 0.65,
      "purpose": "Representative medium model for key scaling point"
    },
    "trainscratch01/d1024_l16_h16": {
      "parameters": "~500M",
      "hidden_size": 1024,
      "num_layers": 16, 
      "num_heads": 16,
      "intermediate_size": 4096,
      "model_size_gb": 1.13,
      "purpose": "Alternative medium size for scaling comparison"
    }
  },
  "usage": {
    "loading_models": {
      "library": "transformers",
      "example": "AutoModelForCausalLM.from_pretrained('ericjm/narrow-data', subfolder='trainscratch01/d768_l12_h12/final_model')"
    },
    "tokenizer": {
      "compatible": "NousResearch/Meta-Llama-3.1-8B",
      "note": "Use this tokenizer for compatibility with all models"
    },
    "training_curves": {
      "location": "trainer_state.json within each final_model directory",
      "description": "Contains step-by-step training history and loss curves"
    }
  },
  "paper_figures": {
    "Figure 6": "LLM training frontiers - uses scaling analysis from these models",
    "Figure 12": "Training run comparison - compares training efficiency across model sizes"
  },
  "technical_details": {
    "model_format": "SafeTensors",
    "precision": "float32",
    "total_upload_size_gb": 1.93,
    "files_per_model": ["model.safetensors", "config.json", "tokenizer.json", "trainer_state.json", "training_args.bin"],
    "excluded_files": ["pruning_mask.pt (5GB each)", "large intermediate checkpoints"],
    "optimization": "Essential final models only for efficient sharing"
  },
  "citation": {
    "bibtex": "@article{michaud2024narrow, title={On the creation of narrow AI: hierarchy and nonlocality of neural network skills}, author={Michaud, Eric and Parker-Sartori, Asher and Tegmark, Max}, journal={arXiv preprint}, year={2024}}"
  }
}