frontend/src/data/sae_features.json

{
  "features": [
    {
      "id": 1,
      "label": "DNA / Genomics",
      "tokens": ["ATCG", "genome", "sequence", "helix", "DNA", "nucleotide", "chromosome", "gene"],
      "activations": [0.95, 0.91, 0.89, 0.88, 0.94, 0.82, 0.79, 0.85],
      "histogram": [0.02, 0.05, 0.08, 0.35, 0.50],
      "monosemantic": true,
      "category": "science"
    },
    {
      "id": 124,
      "label": "HTTP Headers",
      "tokens": ["GET", "POST", "Content-Type", "200 OK", "HTTP", "Authorization", "Accept", "Host"],
      "activations": [0.93, 0.91, 0.88, 0.87, 0.92, 0.84, 0.81, 0.79],
      "histogram": [0.03, 0.06, 0.11, 0.38, 0.42],
      "monosemantic": true,
      "category": "code"
    },
    {
      "id": 331,
      "label": "Arabic Script",
      "tokens": ["\u0627\u0644", "\u0645\u0646", "\u0641\u064a", "\u0639\u0644\u0649", "\u0647\u0630\u0627", "\u0623\u0646", "\u0643\u0627\u0646", "\u0628\u064a\u0646"],
      "activations": [0.92, 0.89, 0.87, 0.85, 0.84, 0.83, 0.81, 0.79],
      "histogram": [0.01, 0.04, 0.09, 0.32, 0.54],
      "monosemantic": true,
      "category": "language"
    },
    {
      "id": 512,
      "label": "Legal Language",
      "tokens": ["plaintiff", "defendant", "hereby", "pursuant", "jurisdiction", "statute", "counsel", "verdict"],
      "activations": [0.94, 0.91, 0.89, 0.87, 0.85, 0.83, 0.80, 0.78],
      "histogram": [0.02, 0.07, 0.12, 0.36, 0.43],
      "monosemantic": true,
      "category": "domain"
    },
    {
      "id": 847,
      "label": "Python Code",
      "tokens": ["def", "class", "import", "return", "self", "__init__", "print", "for"],
      "activations": [0.96, 0.93, 0.91, 0.90, 0.88, 0.86, 0.84, 0.82],
      "histogram": [0.01, 0.03, 0.07, 0.29, 0.60],
      "monosemantic": true,
      "category": "code"
    },
    {
      "id": 1204,
      "label": "Nutritional Labels",
      "tokens": ["calories", "protein", "fat", "mg", "serving", "sodium", "fiber", "carbs"],
      "activations": [0.91, 0.88, 0.86, 0.85, 0.83, 0.81, 0.79, 0.77],
      "histogram": [0.03, 0.08, 0.14, 0.37, 0.38],
      "monosemantic": true,
      "category": "domain"
    },
    {
      "id": 1580,
      "label": "Chess Notation",
      "tokens": ["Nf3", "e4", "O-O", "checkmate", "Qd5", "Bxc6", "Rook", "pawn"],
      "activations": [0.93, 0.90, 0.88, 0.87, 0.85, 0.83, 0.80, 0.78],
      "histogram": [0.02, 0.05, 0.10, 0.33, 0.50],
      "monosemantic": true,
      "category": "domain"
    },
    {
      "id": 1923,
      "label": "French Language",
      "tokens": ["le", "la", "les", "des", "une", "dans", "pour", "avec"],
      "activations": [0.94, 0.92, 0.90, 0.88, 0.86, 0.84, 0.82, 0.80],
      "histogram": [0.01, 0.04, 0.08, 0.31, 0.56],
      "monosemantic": true,
      "category": "language"
    },
    {
      "id": 2341,
      "label": "Markdown Headers",
      "tokens": ["##", "###", "**bold**", "[link]", "---", "> quote", "```", "* list"],
      "activations": [0.92, 0.89, 0.87, 0.85, 0.83, 0.81, 0.80, 0.78],
      "histogram": [0.02, 0.06, 0.11, 0.34, 0.47],
      "monosemantic": true,
      "category": "code"
    },
    {
      "id": 2788,
      "label": "German Language",
      "tokens": ["der", "die", "das", "und", "ist", "nicht", "ein", "mit"],
      "activations": [0.93, 0.91, 0.89, 0.87, 0.85, 0.83, 0.81, 0.79],
      "histogram": [0.01, 0.05, 0.09, 0.32, 0.53],
      "monosemantic": true,
      "category": "language"
    },
    {
      "id": 3102,
      "label": "Medical Terms",
      "tokens": ["diagnosis", "treatment", "patient", "symptoms", "therapy", "clinical", "dosage", "prognosis"],
      "activations": [0.94, 0.92, 0.90, 0.88, 0.86, 0.84, 0.82, 0.80],
      "histogram": [0.02, 0.05, 0.10, 0.35, 0.48],
      "monosemantic": true,
      "category": "domain"
    },
    {
      "id": 3847,
      "label": "Shakespeare",
      "tokens": ["thee", "thou", "dost", "hath", "wherefore", "prithee", "forsooth", "methinks"],
      "activations": [0.95, 0.93, 0.91, 0.89, 0.87, 0.85, 0.83, 0.81],
      "histogram": [0.01, 0.04, 0.08, 0.30, 0.57],
      "monosemantic": true,
      "category": "language"
    }
  ],
  "training": {
    "steps": [0, 10000, 20000, 30000, 40000, 50000, 60000, 70000, 80000, 90000, 100000, 110000, 120000, 130000, 140000, 150000, 160000, 170000, 180000, 190000, 200000],
    "total_loss": [0.82, 0.61, 0.48, 0.39, 0.34, 0.30, 0.27, 0.25, 0.24, 0.23, 0.22, 0.21, 0.21, 0.20, 0.20, 0.19, 0.19, 0.19, 0.18, 0.18, 0.18],
    "recon_loss": [0.62, 0.44, 0.34, 0.27, 0.23, 0.20, 0.18, 0.17, 0.16, 0.15, 0.15, 0.14, 0.14, 0.13, 0.13, 0.13, 0.12, 0.12, 0.12, 0.12, 0.12],
    "annotations": [
      { "step": 50000, "label": "First neuron resampling" },
      { "step": 150000, "label": "L1 coefficient sweep" }
    ]
  },
  "meta": {
    "model": "GPT-2 Small",
    "layer": 6,
    "d_model": 512,
    "expansion": 8,
    "d_hidden": 4096,
    "l1_coeff": 0.001,
    "alive_pct": 93,
    "interpretable_pct": 70,
    "recon_loss": 0.12,
    "paper": "Bricken, Templeton et al. (Anthropic, 2023)",
    "title": "Towards Monosemanticity"
  }
}