data/features.json

{
  "_meta": {
    "model": "google/gemma-3-4b-it",
    "sae_release": "gemma-scope-2-4b-it-res",
    "sae_id_template": "layer_{layer}_width_16k_l0_medium",
    "sae_width": 16384,
    "hook_point_template": "blocks.{layer}.hook_resid_post",
    "neuronpedia_template": "https://www.neuronpedia.org/gemma-3-4b-it/{layer}-gemmascope-2-res-16k/{feat}",
    "clamp_rule": "clamp_target = 2 * max_act (max_act = peak activation on the 426-noun feature-detection corpus)",
    "source": "submission/sae_steering/feature_analysis/interpretations/resid_post/"
  },
  "features": [
    {
      "key": "L22.F193",
      "layer": 22,
      "feat_id": 193,
      "label": "Edibles / food",
      "tldr": "Edible foods",
      "dimensions": [
        "taste_smell"
      ],
      "primary_axis": "taste_smell",
      "sae_r": 0.559,
      "llm_r": 0.065,
      "ceiling": 0.887,
      "max_act": 1113.11,
      "clamp_target": 2226.23,
      "n_active": 43,
      "neuronpedia": "https://www.neuronpedia.org/gemma-3-4b-it/22-gemmascope-2-res-16k/193",
      "top_nouns": [
        [
          "ham",
          1113.1,
          "Sensorimotor/Gustatory"
        ],
        [
          "sandwich",
          1024.1,
          "Sensorimotor/Mouth"
        ],
        [
          "toast",
          980.4,
          "Sensorimotor/Gustatory"
        ],
        [
          "zucchini",
          948.2,
          "Sensorimotor/Gustatory"
        ],
        [
          "muffin",
          866.4,
          "Sensorimotor/Mouth"
        ],
        [
          "cocktail",
          821.7,
          "Sensorimotor/Gustatory"
        ],
        [
          "spinach",
          795.5,
          "Sensorimotor/Gustatory"
        ],
        [
          "salty",
          788.0,
          "Sensorimotor/Gustatory"
        ],
        [
          "kiwi",
          712.9,
          "Sensorimotor/Gustatory"
        ],
        [
          "melon",
          693.4,
          "Sensorimotor/Gustatory"
        ]
      ],
      "corr": [
        [
          "taste_smell",
          0.559,
          0.065
        ],
        [
          "color",
          0.33,
          0.033
        ],
        [
          "tactile",
          0.302,
          -0.088
        ],
        [
          "thought",
          -0.245,
          0.634
        ],
        [
          "visual_form",
          0.241,
          -0.128
        ],
        [
          "moral",
          -0.21,
          0.574
        ],
        [
          "time",
          -0.206,
          0.366
        ],
        [
          "social",
          -0.199,
          0.499
        ],
        [
          "emotion",
          -0.125,
          0.439
        ],
        [
          "space",
          -0.105,
          0.333
        ],
        [
          "quantity",
          -0.104,
          0.403
        ],
        [
          "motion_self",
          -0.094,
          0.418
        ],
        [
          "auditory",
          -0.066,
          -0.007
        ],
        [
          "polarity",
          -0.007,
          0.512
        ]
      ],
      "interpretation": "Edibles / food detector. All top corpus nouns are food-related and Sensorimotor/Gustatory or Mouth: ham, sandwich, toast, zucchini, muffin, cocktail, spinach, salty, kiwi, melon. Highly selective: only 17/426 corpus nouns fire. Narrow but clean signature: taste_smell +0.56 dominates, with smaller bleed-through to color +0.33 / tactile +0.30 / visual_form +0.24.",
      "steering_verdict": "Steers generations toward food vocabulary, improving alignment with taste_smell; other axes change little because the feature fires on only a small subset of nouns."
    },
    {
      "key": "L17.F261",
      "layer": 17,
      "feat_id": 261,
      "label": "BodyFood (embodied noun)",
      "tldr": "Body parts and food",
      "dimensions": [
        "visual_form",
        "tactile",
        "color"
      ],
      "primary_axis": "visual_form",
      "sae_r": 0.584,
      "llm_r": -0.128,
      "ceiling": 0.953,
      "max_act": 608.7,
      "clamp_target": 1217.4,
      "n_active": 460,
      "neuronpedia": "https://www.neuronpedia.org/gemma-3-4b-it/17-gemmascope-2-res-16k/261",
      "top_nouns": [
        [
          "appendectomy",
          608.7,
          "Sensorimotor/Torso"
        ],
        [
          "kidney",
          570.1,
          "Sensorimotor/Torso"
        ],
        [
          "hairline",
          552.7,
          "Sensorimotor/Head"
        ],
        [
          "abdomen",
          466.6,
          "Sensorimotor/Torso"
        ],
        [
          "ham",
          442.7,
          "Sensorimotor/Gustatory"
        ],
        [
          "forearm",
          441.4,
          "Sensorimotor/Hand_arm"
        ],
        [
          "toast",
          435.4,
          "Sensorimotor/Gustatory"
        ],
        [
          "kiwi",
          414.4,
          "Sensorimotor/Gustatory"
        ],
        [
          "arm",
          405.5,
          "Sensorimotor/Hand_arm"
        ],
        [
          "stomach",
          397.6,
          "Sensorimotor/Torso"
        ]
      ],
      "corr": [
        [
          "visual_form",
          0.584,
          -0.128
        ],
        [
          "tactile",
          0.581,
          -0.088
        ],
        [
          "taste_smell",
          0.528,
          0.065
        ],
        [
          "social",
          -0.517,
          0.499
        ],
        [
          "thought",
          -0.503,
          0.634
        ],
        [
          "color",
          0.5,
          0.033
        ],
        [
          "moral",
          -0.479,
          0.574
        ],
        [
          "time",
          -0.424,
          0.366
        ],
        [
          "emotion",
          -0.402,
          0.439
        ],
        [
          "motion_self",
          -0.34,
          0.418
        ],
        [
          "polarity",
          -0.26,
          0.512
        ],
        [
          "quantity",
          -0.21,
          0.403
        ],
        [
          "auditory",
          0.142,
          -0.007
        ],
        [
          "space",
          -0.001,
          0.333
        ]
      ],
      "interpretation": "Embodied-noun detector. Top nouns are body parts (appendectomy, kidney, abdomen, forearm, arm, stomach) and edibles (ham, toast, kiwi); all Sensorimotor. The textbook concrete-vs-abstract pattern: positive on every sensorimotor rated dim (visual_form +0.58, tactile +0.58, taste_smell +0.53, color +0.50) and strongly negative on social (-0.52), thought (-0.50), moral (-0.48), because abstract words score high on those dims and the feature does not fire on them.",
      "steering_verdict": "Steers generations toward body parts and food, improving alignment with sensorimotor axes (visual_form, tactile, color, taste_smell); abstract, social, moral, and thought properties are reduced."
    },
    {
      "key": "L22.F472",
      "layer": 22,
      "feat_id": 472,
      "label": "High-magnitude events",
      "tldr": "Consequential high-stakes events",
      "dimensions": [
        "time",
        "quantity"
      ],
      "primary_axis": "time",
      "sae_r": 0.498,
      "llm_r": 0.366,
      "ceiling": 0.883,
      "max_act": 1864.8,
      "clamp_target": 3729.6,
      "n_active": 615,
      "neuronpedia": "https://www.neuronpedia.org/gemma-3-4b-it/22-gemmascope-2-res-16k/472",
      "top_nouns": [
        [
          "scientist",
          1864.8,
          "Internal/Glasgow_pure"
        ],
        [
          "tyrant",
          1549.9,
          "Internal/Glasgow_pure"
        ],
        [
          "whore",
          1421.9,
          "Internal/Glasgow_pure"
        ],
        [
          "earthquake",
          1319.3,
          "Internal/Glasgow_pure"
        ],
        [
          "jail",
          1293.3,
          "Internal/Glasgow_pure"
        ],
        [
          "slave",
          1276.9,
          "Internal/Glasgow_pure"
        ],
        [
          "bomb",
          1229.6,
          "Internal/Glasgow_pure"
        ],
        [
          "prisoner",
          1069.2,
          "Internal/Secondary_Social"
        ],
        [
          "hell",
          1013.3,
          "Internal/Glasgow_pure"
        ],
        [
          "war",
          1006.2,
          "Internal/Glasgow_pure"
        ]
      ],
      "corr": [
        [
          "visual_form",
          -0.619,
          -0.128
        ],
        [
          "thought",
          0.523,
          0.634
        ],
        [
          "tactile",
          -0.5,
          -0.088
        ],
        [
          "time",
          0.498,
          0.366
        ],
        [
          "color",
          -0.472,
          0.033
        ],
        [
          "auditory",
          -0.45,
          -0.007
        ],
        [
          "taste_smell",
          -0.45,
          0.065
        ],
        [
          "moral",
          0.428,
          0.574
        ],
        [
          "quantity",
          0.424,
          0.403
        ],
        [
          "motion_self",
          0.33,
          0.418
        ],
        [
          "social",
          0.217,
          0.499
        ],
        [
          "emotion",
          0.2,
          0.439
        ],
        [
          "space",
          0.156,
          0.333
        ],
        [
          "polarity",
          0.133,
          0.512
        ]
      ],
      "interpretation": "High-magnitude / consequential abstract-event detector. Top nouns mix confinement and harm (jail, hell, prisoner, war), destructive forces (earthquake, bomb), oppressive actors (tyrant, slave), and high-stakes roles (scientist). Broadly active (615/619 rated words). Positive on thought (+0.52), time (+0.50), moral (+0.43), quantity (+0.42); strongly negative on every perceptual axis.",
      "steering_verdict": "Steers generations toward high-impact abstract events (war, jail, bomb, earthquake), improving alignment with time; perceptual properties are reduced."
    },
    {
      "key": "L29.F407",
      "layer": 29,
      "feat_id": 407,
      "label": "Situation (neg. leaning)",
      "tldr": "Difficult situations and circumstances",
      "dimensions": [
        "emotion"
      ],
      "primary_axis": "emotion",
      "sae_r": 0.577,
      "llm_r": 0.439,
      "ceiling": 0.879,
      "max_act": 2050.3,
      "clamp_target": 4100.6,
      "n_active": 443,
      "neuronpedia": "https://www.neuronpedia.org/gemma-3-4b-it/29-gemmascope-2-res-16k/407",
      "top_nouns": [
        [
          "opportunity",
          2050.3,
          "Internal/Glasgow_pure"
        ],
        [
          "worry",
          1857.0,
          "Sensorimotor/Head"
        ],
        [
          "circumstances",
          1825.4,
          "Abstract"
        ],
        [
          "complication",
          1650.8,
          "Abstract"
        ],
        [
          "setback",
          1630.7,
          "Abstract"
        ],
        [
          "conflict",
          1481.0,
          "Social"
        ],
        [
          "pressure",
          1420.4,
          "Internal/Glasgow_pure"
        ],
        [
          "debacle",
          1249.4,
          "Internal/Glasgow_pure"
        ],
        [
          "irrevocableness",
          1186.0,
          "Abstract"
        ],
        [
          "tragic",
          1177.3,
          "Internal/Glasgow_pure"
        ]
      ],
      "corr": [
        [
          "emotion",
          0.577,
          0.439
        ],
        [
          "moral",
          0.487,
          0.574
        ],
        [
          "social",
          0.459,
          0.499
        ],
        [
          "thought",
          0.458,
          0.634
        ],
        [
          "polarity",
          0.406,
          0.512
        ],
        [
          "visual_form",
          -0.403,
          -0.128
        ],
        [
          "motion_self",
          0.353,
          0.418
        ],
        [
          "tactile",
          -0.291,
          -0.088
        ],
        [
          "color",
          -0.278,
          0.033
        ],
        [
          "taste_smell",
          -0.27,
          0.065
        ],
        [
          "time",
          0.251,
          0.366
        ],
        [
          "auditory",
          -0.092,
          -0.007
        ],
        [
          "space",
          -0.049,
          0.333
        ],
        [
          "quantity",
          0.046,
          0.403
        ]
      ],
      "interpretation": "Negative-leaning situation / circumstance feature. Top nouns name situations a person finds themselves in: opportunity, worry, circumstances, complication, setback, conflict, pressure, debacle, tragic. Fires on 443/619 rated words. Uniformly positive across Internal-Social axes: emotion +0.58, moral +0.49, social +0.46, thought +0.46, polarity +0.41.",
      "steering_verdict": "Steers generations toward situational and evaluative content (opportunity, conflict, setback, pressure), improving alignment with emotion; perceptual properties are reduced."
    },
    {
      "key": "L29.F132",
      "layer": 29,
      "feat_id": 132,
      "label": "Charged event",
      "tldr": "Charged abstract events",
      "dimensions": [
        "motion_self"
      ],
      "primary_axis": "motion_self",
      "sae_r": 0.479,
      "llm_r": 0.418,
      "ceiling": 0.858,
      "max_act": 4005.7,
      "clamp_target": 8011.39,
      "n_active": 341,
      "neuronpedia": "https://www.neuronpedia.org/gemma-3-4b-it/29-gemmascope-2-res-16k/132",
      "top_nouns": [
        [
          "marvel",
          4005.7,
          "Internal/Glasgow_pure"
        ],
        [
          "war",
          4002.3,
          "Internal/Glasgow_pure"
        ],
        [
          "prick",
          3107.7,
          "Internal/Glasgow_pure"
        ],
        [
          "inexorableness",
          3051.3,
          "Abstract"
        ],
        [
          "progression",
          2952.5,
          "Internal/Glasgow_pure"
        ],
        [
          "truth",
          2630.0,
          "Internal/Glasgow_pure"
        ],
        [
          "avalanche",
          2489.3,
          "Internal/Glasgow_pure"
        ],
        [
          "jail",
          2465.5,
          "Internal/Glasgow_pure"
        ],
        [
          "skill",
          2141.3,
          "Internal/Glasgow_pure"
        ],
        [
          "talent",
          2114.4,
          "Internal/Glasgow_pure"
        ]
      ],
      "corr": [
        [
          "thought",
          0.555,
          0.634
        ],
        [
          "emotion",
          0.525,
          0.439
        ],
        [
          "moral",
          0.493,
          0.574
        ],
        [
          "motion_self",
          0.479,
          0.418
        ],
        [
          "visual_form",
          -0.469,
          -0.128
        ],
        [
          "taste_smell",
          -0.403,
          0.065
        ],
        [
          "tactile",
          -0.376,
          -0.088
        ],
        [
          "polarity",
          0.292,
          0.512
        ],
        [
          "color",
          -0.264,
          0.033
        ],
        [
          "social",
          0.254,
          0.499
        ],
        [
          "auditory",
          -0.217,
          -0.007
        ],
        [
          "time",
          0.213,
          0.366
        ],
        [
          "quantity",
          0.138,
          0.403
        ],
        [
          "space",
          -0.019,
          0.333
        ]
      ],
      "interpretation": "Abstract-magnitude / charged-event detector. Top nouns mix large-scale events (war, avalanche, jail), high-impact positives (marvel, truth, skill, talent), and abstract progression terms. Fires on 341/619 rated words. Positive on the abstract/internal cluster (thought +0.56, emotion +0.53, moral +0.49, motion_self +0.48) and negative on all five perceptual axes.",
      "steering_verdict": "Steers generations toward high-impact abstract and affective content (war, marvel, talent, truth), improving alignment with motion_self and emotion; perceptual properties are reduced."
    },
    {
      "key": "L29.F1574",
      "layer": 29,
      "feat_id": 1574,
      "label": "Pure auditory",
      "tldr": "Sounds and noises",
      "dimensions": [
        "auditory"
      ],
      "primary_axis": "auditory",
      "sae_r": 0.44,
      "llm_r": -0.007,
      "ceiling": 0.92,
      "max_act": 3541.1,
      "clamp_target": 7082.2,
      "n_active": 64,
      "neuronpedia": "https://www.neuronpedia.org/gemma-3-4b-it/29-gemmascope-2-res-16k/1574",
      "top_nouns": [
        [
          "murmur",
          3541.1,
          "Sensorimotor/Auditory"
        ],
        [
          "plea",
          1814.8,
          "Sensorimotor/Auditory"
        ],
        [
          "pronounce",
          954.6,
          "Sensorimotor/Mouth"
        ],
        [
          "detonate",
          952.9,
          "Internal/Glasgow_pure"
        ],
        [
          "duet",
          910.2,
          "Sensorimotor/Auditory"
        ],
        [
          "drums",
          705.2,
          "Sensorimotor/Auditory"
        ],
        [
          "rhyme",
          693.4,
          "Sensorimotor/Auditory"
        ]
      ],
      "corr": [
        [
          "auditory",
          0.44,
          -0.007
        ],
        [
          "moral",
          -0.178,
          0.574
        ],
        [
          "thought",
          -0.142,
          0.634
        ],
        [
          "social",
          -0.085,
          0.499
        ],
        [
          "polarity",
          -0.085,
          0.512
        ],
        [
          "time",
          -0.078,
          0.366
        ],
        [
          "quantity",
          -0.07,
          0.403
        ],
        [
          "visual_form",
          0.057,
          -0.128
        ],
        [
          "tactile",
          0.053,
          -0.088
        ],
        [
          "color",
          0.042,
          0.033
        ],
        [
          "emotion",
          -0.038,
          0.439
        ],
        [
          "motion_self",
          -0.021,
          0.418
        ],
        [
          "taste_smell",
          0.012,
          0.065
        ],
        [
          "space",
          0.01,
          0.333
        ]
      ],
      "interpretation": "Pure auditory detector. Only 7 corpus nouns fire, dominated by Sensorimotor/Auditory: murmur, plea, duet, drums, rhyme; plus pronounce and detonate. Single-axis signature: auditory +0.44 (LLM r is essentially zero); every other dim near zero. The cleanest single-axis auditory feature in the resid_post 4B set, but extreme selectivity means it only fires on words with explicit auditory content.",
      "steering_verdict": "Steers generations toward auditory vocabulary (murmur, drums, duet, rhyme), improving alignment with auditory; effect on other axes is small."
    },
    {
      "key": "L22.F166",
      "layer": 22,
      "feat_id": 166,
      "label": "Family / kinship",
      "tldr": "Family and kinship",
      "dimensions": [
        "social"
      ],
      "primary_axis": "social",
      "sae_r": 0.401,
      "llm_r": 0.499,
      "ceiling": 0.895,
      "max_act": 1290.5,
      "clamp_target": 2581.0,
      "n_active": 31,
      "neuronpedia": "https://www.neuronpedia.org/gemma-3-4b-it/22-gemmascope-2-res-16k/166",
      "top_nouns": [
        [
          "mommy",
          1290.5,
          "Social"
        ],
        [
          "grandson",
          1228.5,
          "Social"
        ],
        [
          "grandchild",
          1087.4,
          "Social"
        ],
        [
          "grandpa",
          973.2,
          "Social"
        ],
        [
          "husband",
          950.6,
          "Social"
        ],
        [
          "motherhood",
          926.6,
          "Social"
        ],
        [
          "courtship",
          860.2,
          "Social"
        ],
        [
          "cousin",
          846.2,
          "Social"
        ],
        [
          "mates",
          835.4,
          "Internal/Glasgow_pure"
        ],
        [
          "paternity",
          780.4,
          "Social"
        ]
      ],
      "corr": [
        [
          "social",
          0.401,
          0.499
        ],
        [
          "emotion",
          0.286,
          0.439
        ],
        [
          "polarity",
          0.275,
          0.512
        ],
        [
          "auditory",
          0.105,
          -0.007
        ],
        [
          "moral",
          0.094,
          0.574
        ],
        [
          "quantity",
          -0.08,
          0.403
        ],
        [
          "thought",
          0.06,
          0.634
        ],
        [
          "tactile",
          0.055,
          -0.088
        ],
        [
          "taste_smell",
          0.052,
          0.065
        ],
        [
          "motion_self",
          0.049,
          0.418
        ],
        [
          "space",
          -0.046,
          0.333
        ],
        [
          "time",
          0.032,
          0.366
        ],
        [
          "visual_form",
          0.022,
          -0.128
        ],
        [
          "color",
          -0.019,
          0.033
        ]
      ],
      "interpretation": "Family and kinship-relations detector. Top nouns are kin and relationship terms: mommy, grandson, grandchild, grandpa, husband, motherhood, courtship, cousin, paternity. Fires on 31 rated words. The social dimension is its strongest correlate (+0.40); the bare model already rates social highly, so this feature is the paper's representative social feature on Neuronpedia rather than a dormant one.",
      "steering_verdict": "Steers generations toward family and kinship relations, improving alignment with social."
    },
    {
      "key": "L29.F472",
      "layer": 29,
      "feat_id": 472,
      "label": "Affect / emotional states",
      "tldr": "Emotional states",
      "dimensions": [
        "emotion"
      ],
      "primary_axis": "emotion",
      "sae_r": 0.563,
      "llm_r": 0.439,
      "ceiling": 0.879,
      "max_act": 4628.1,
      "clamp_target": 9256.2,
      "n_active": 126,
      "neuronpedia": "https://www.neuronpedia.org/gemma-3-4b-it/29-gemmascope-2-res-16k/472",
      "top_nouns": [
        [
          "desperation",
          4628.1,
          "Abstract"
        ],
        [
          "repugnance",
          4609.6,
          "Abstract"
        ],
        [
          "expectance",
          3874.4,
          "Abstract"
        ],
        [
          "boredom",
          3563.5,
          "Internal/Plutchik"
        ],
        [
          "disinclination",
          3271.6,
          "Abstract"
        ],
        [
          "mystification",
          3118.6,
          "Abstract"
        ],
        [
          "pensiveness",
          3065.8,
          "Internal/Plutchik"
        ],
        [
          "dubiousness",
          2997.6,
          "Abstract"
        ],
        [
          "mistrustfulness",
          2951.0,
          "Abstract"
        ],
        [
          "elation",
          2931.8,
          "Internal/Glasgow_pure"
        ]
      ],
      "corr": [
        [
          "emotion",
          0.563,
          0.439
        ],
        [
          "polarity",
          0.355,
          0.512
        ],
        [
          "moral",
          0.335,
          0.574
        ],
        [
          "thought",
          0.323,
          0.634
        ],
        [
          "motion_self",
          0.299,
          0.418
        ],
        [
          "social",
          0.274,
          0.499
        ],
        [
          "visual_form",
          -0.229,
          -0.128
        ],
        [
          "taste_smell",
          -0.154,
          0.065
        ],
        [
          "tactile",
          -0.141,
          -0.088
        ],
        [
          "color",
          -0.127,
          0.033
        ],
        [
          "quantity",
          -0.126,
          0.403
        ],
        [
          "space",
          -0.077,
          0.333
        ],
        [
          "auditory",
          -0.048,
          -0.007
        ],
        [
          "time",
          0.041,
          0.366
        ]
      ],
      "interpretation": "Emotional and motivational states across positive and negative affect: desperation, repugnance, boredom, pensiveness, mistrustfulness, elation. Fires on 126 rated words. Positive across the internal/evaluative cluster (emotion +0.56, polarity +0.36, moral +0.33, thought +0.32) and negative on perceptual axes. The paper's representative Internal feature. Note: this is L29.F472, distinct from the L22.F472 high-magnitude-events feature.",
      "steering_verdict": "Steers generations toward emotional and affective states, improving alignment with emotion (internal)."
    },
    {
      "key": "L29.F31",
      "layer": 29,
      "feat_id": 31,
      "label": "Consequential magnitude",
      "tldr": "High-stakes events and figures",
      "dimensions": [
        "time",
        "thought",
        "quantity"
      ],
      "primary_axis": "time",
      "sae_r": 0.487,
      "llm_r": 0.366,
      "ceiling": 0.883,
      "max_act": 3383.0,
      "clamp_target": 6766.0,
      "n_active": 91,
      "neuronpedia": "https://www.neuronpedia.org/gemma-3-4b-it/29-gemmascope-2-res-16k/31",
      "top_nouns": [
        [
          "earthquake",
          3383.0,
          "Internal/Glasgow_pure"
        ],
        [
          "scientist",
          3219.0,
          "Internal/Glasgow_pure"
        ],
        [
          "whore",
          2605.0,
          "Internal/Glasgow_pure"
        ],
        [
          "bomb",
          2408.0,
          "Internal/Glasgow_pure"
        ],
        [
          "tyrant",
          2334.0,
          "Internal/Glasgow_pure"
        ],
        [
          "goddess",
          1965.0,
          "Social"
        ],
        [
          "gem",
          1927.0,
          "Sensorimotor/Visual"
        ],
        [
          "determinist",
          1875.0,
          "Abstract"
        ]
      ],
      "corr": [
        [
          "visual_form",
          -0.554,
          -0.128
        ],
        [
          "tactile",
          -0.494,
          -0.088
        ],
        [
          "time",
          0.487,
          0.366
        ],
        [
          "color",
          -0.482,
          0.033
        ],
        [
          "thought",
          0.471,
          0.634
        ],
        [
          "quantity",
          0.416,
          0.403
        ],
        [
          "taste_smell",
          -0.413,
          0.065
        ],
        [
          "auditory",
          -0.357,
          -0.007
        ],
        [
          "moral",
          0.355,
          0.574
        ],
        [
          "motion_self",
          0.291,
          0.418
        ],
        [
          "social",
          0.235,
          0.499
        ],
        [
          "space",
          0.221,
          0.333
        ],
        [
          "emotion",
          0.112,
          0.439
        ],
        [
          "polarity",
          0.112,
          0.512
        ]
      ],
      "interpretation": "High-magnitude / consequential entity detector: earthquake, scientist, bomb, tyrant, goddess. Fires on 91 rated words. Positive on time (+0.49), thought (+0.47), quantity (+0.42), moral (+0.35); strongly negative on perceptual axes. A second, sharper variant of the magnitude pattern alongside L22.F472.",
      "steering_verdict": "Steers generations toward high-stakes events and consequential figures, improving alignment with time, thought, and quantity."
    },
    {
      "key": "L22.F451",
      "layer": 22,
      "feat_id": 451,
      "label": "Spatial extent",
      "tldr": "Space, extent, scope",
      "dimensions": [
        "space"
      ],
      "primary_axis": "space",
      "sae_r": 0.335,
      "llm_r": 0.333,
      "ceiling": 0.808,
      "max_act": 1242.0,
      "clamp_target": 2484.0,
      "n_active": 16,
      "neuronpedia": "https://www.neuronpedia.org/gemma-3-4b-it/22-gemmascope-2-res-16k/451",
      "top_nouns": [
        [
          "texture",
          1242.0,
          "Sensorimotor/Haptic"
        ],
        [
          "admissibleness",
          881.0,
          "Abstract"
        ],
        [
          "allowableness",
          796.0,
          "Abstract"
        ],
        [
          "validity",
          792.0,
          "Abstract"
        ],
        [
          "religion",
          752.0,
          "Abstract"
        ],
        [
          "formalness",
          554.0,
          "Abstract"
        ],
        [
          "relativity",
          479.0,
          "Abstract"
        ],
        [
          "extent",
          434.0,
          "Abstract"
        ]
      ],
      "corr": [
        [
          "space",
          0.335,
          0.333
        ],
        [
          "emotion",
          -0.191,
          0.439
        ],
        [
          "time",
          0.185,
          0.366
        ],
        [
          "quantity",
          0.151,
          0.403
        ],
        [
          "moral",
          -0.095,
          0.574
        ],
        [
          "motion_self",
          -0.066,
          0.418
        ],
        [
          "thought",
          -0.043,
          0.634
        ],
        [
          "polarity",
          -0.041,
          0.512
        ],
        [
          "color",
          -0.036,
          0.033
        ],
        [
          "visual_form",
          0.033,
          -0.128
        ],
        [
          "auditory",
          0.031,
          -0.007
        ],
        [
          "tactile",
          -0.029,
          -0.088
        ],
        [
          "taste_smell",
          -0.019,
          0.065
        ],
        [
          "social",
          -0.014,
          0.499
        ]
      ],
      "interpretation": "Spatial-extent feature: its strongest correlate is the space dimension (+0.34), the only such feature in this set. Fires on a small set of words (extent, relativity, validity, texture). The bare model already rates space moderately, so the margin is small.",
      "steering_verdict": "Weakly steers toward spatial extent and scope; the effect is subtle because the feature fires on few words."
    },
    {
      "key": "L29.F353",
      "layer": 29,
      "feat_id": 353,
      "label": "Control feature (uncorrelated)",
      "tldr": "Control: no clear dimension",
      "dimensions": [],
      "primary_axis": null,
      "sae_r": 0.143,
      "llm_r": 0.499,
      "ceiling": 0.0,
      "max_act": 4934.0,
      "clamp_target": 9868.0,
      "n_active": 273,
      "is_control": true,
      "neuronpedia": "https://www.neuronpedia.org/gemma-3-4b-it/29-gemmascope-2-res-16k/353",
      "top_nouns": [
        [
          "magnificence",
          4934.0,
          "Abstract"
        ],
        [
          "incongruity",
          4436.0,
          "Abstract"
        ],
        [
          "aroma",
          4000.0,
          "Sensorimotor/Olfactory"
        ],
        [
          "synthesis",
          3971.0,
          "Abstract"
        ],
        [
          "cocktail",
          3556.0,
          "Sensorimotor/Gustatory"
        ],
        [
          "dynasty",
          3466.0,
          "Social"
        ],
        [
          "heroine",
          3416.0,
          "Social"
        ],
        [
          "war",
          3372.0,
          "Internal/Glasgow_pure"
        ]
      ],
      "corr": [
        [
          "social",
          0.143,
          0.499
        ],
        [
          "auditory",
          0.142,
          -0.007
        ],
        [
          "time",
          0.121,
          0.366
        ],
        [
          "polarity",
          0.099,
          0.512
        ],
        [
          "quantity",
          0.093,
          0.403
        ],
        [
          "space",
          0.08,
          0.333
        ],
        [
          "color",
          0.071,
          0.033
        ],
        [
          "motion_self",
          0.061,
          0.418
        ],
        [
          "thought",
          0.057,
          0.634
        ],
        [
          "emotion",
          0.049,
          0.439
        ],
        [
          "moral",
          -0.037,
          0.574
        ],
        [
          "taste_smell",
          -0.021,
          0.065
        ],
        [
          "tactile",
          -0.008,
          -0.088
        ],
        [
          "visual_form",
          0.006,
          -0.128
        ]
      ],
      "interpretation": "Control feature: no rating dimension correlates above |0.14|. It fires broadly on a semantically mixed set (magnificence, incongruity, aroma, synthesis, cocktail, dynasty), with no coherent grounding theme. Included as a contrast to the grounding-aligned features.",
      "steering_verdict": "Clamping this feature should NOT push generations toward any particular grounding dimension; use it as a control to see what an uncorrelated feature does."
    },
    {
      "key": "L17.F456",
      "layer": 17,
      "feat_id": 456,
      "label": "Control feature (uncorrelated)",
      "tldr": "Control: no clear dimension",
      "dimensions": [],
      "primary_axis": null,
      "sae_r": 0.027,
      "llm_r": -0.088,
      "ceiling": 0.0,
      "max_act": 1523.0,
      "clamp_target": 3046.0,
      "n_active": 96,
      "is_control": true,
      "neuronpedia": "https://www.neuronpedia.org/gemma-3-4b-it/17-gemmascope-2-res-16k/456",
      "top_nouns": [
        [
          "dope",
          1523.0,
          "Sensorimotor/Olfactory"
        ],
        [
          "existentialism",
          1108.0,
          "Abstract"
        ],
        [
          "constipation",
          1052.0,
          "Internal/Glasgow_pure"
        ],
        [
          "objectivism",
          943.0,
          "Abstract"
        ],
        [
          "diabolism",
          854.0,
          "Abstract"
        ],
        [
          "zucchini",
          798.0,
          "Sensorimotor/Gustatory"
        ],
        [
          "shampoo",
          699.0,
          "Sensorimotor/Olfactory"
        ],
        [
          "alpha",
          694.0,
          "Abstract"
        ]
      ],
      "corr": [
        [
          "quantity",
          -0.168,
          0.403
        ],
        [
          "auditory",
          -0.159,
          -0.007
        ],
        [
          "space",
          -0.15,
          0.333
        ],
        [
          "polarity",
          -0.134,
          0.512
        ],
        [
          "time",
          -0.119,
          0.366
        ],
        [
          "social",
          -0.057,
          0.499
        ],
        [
          "motion_self",
          -0.046,
          0.418
        ],
        [
          "emotion",
          -0.035,
          0.439
        ],
        [
          "tactile",
          0.027,
          -0.088
        ],
        [
          "thought",
          -0.019,
          0.634
        ],
        [
          "moral",
          -0.017,
          0.574
        ],
        [
          "taste_smell",
          -0.017,
          0.065
        ],
        [
          "visual_form",
          -0.004,
          -0.128
        ],
        [
          "color",
          0.001,
          0.033
        ]
      ],
      "interpretation": "Control feature: all 14 rating dimensions correlate below |0.17|. Top words are an incoherent mix (dope, existentialism, constipation, zucchini, shampoo), with no grounding theme. Included as a contrast to the grounding-aligned features.",
      "steering_verdict": "Clamping this feature should NOT push generations toward any grounding dimension; use it as a control."
    },
    {
      "key": "L29.F170",
      "layer": 29,
      "feat_id": 170,
      "label": "Shape and handling",
      "tldr": "Shapes and handling",
      "dimensions": [
        "visual_form"
      ],
      "primary_axis": "visual_form",
      "sae_r": 0.407,
      "llm_r": -0.128,
      "ceiling": 0.953,
      "max_act": 3293.5,
      "clamp_target": 6587.0,
      "n_active": 28,
      "neuronpedia": "https://www.neuronpedia.org/gemma-3-4b-it/29-gemmascope-2-res-16k/170",
      "top_nouns": [
        [
          "shape",
          3293.5,
          "Sensorimotor/Visual"
        ],
        [
          "bumpy",
          2366.6,
          "Sensorimotor/Haptic"
        ],
        [
          "handle",
          2205.5,
          "Sensorimotor/Hand_arm"
        ],
        [
          "hairline",
          1635.4,
          "Sensorimotor/Head"
        ],
        [
          "squeeze",
          1525.8,
          "Sensorimotor/Haptic"
        ],
        [
          "tug",
          1484.4,
          "Sensorimotor/Hand_arm"
        ],
        [
          "coarse",
          1181.4,
          "Sensorimotor/Haptic"
        ],
        [
          "thread",
          1150.9,
          "Sensorimotor/Haptic"
        ],
        [
          "throw",
          1136.6,
          "Sensorimotor/Hand_arm"
        ],
        [
          "pedal",
          1074.9,
          "Sensorimotor/Foot_leg"
        ]
      ],
      "corr": [
        [
          "social",
          -0.481,
          0.499
        ],
        [
          "thought",
          -0.428,
          0.634
        ],
        [
          "moral",
          -0.412,
          0.574
        ],
        [
          "visual_form",
          0.407,
          -0.128
        ],
        [
          "emotion",
          -0.405,
          0.439
        ],
        [
          "polarity",
          -0.39,
          0.512
        ],
        [
          "motion_self",
          -0.311,
          0.418
        ],
        [
          "tactile",
          0.304,
          -0.088
        ],
        [
          "color",
          0.281,
          0.033
        ],
        [
          "time",
          -0.182,
          0.366
        ],
        [
          "taste_smell",
          0.108,
          0.065
        ],
        [
          "quantity",
          0.099,
          0.403
        ],
        [
          "auditory",
          -0.051,
          -0.007
        ],
        [
          "space",
          0.019,
          0.333
        ]
      ],
      "interpretation": "Visual form and physical handling detector: shape, bumpy, handle, squeeze, tug, coarse, thread, throw, pedal. Fires on concrete words that have a graspable shape or texture. Positive on visual_form (+0.41) and tactile (+0.30); strongly negative on the abstract/social axes.",
      "steering_verdict": "Steers generations toward shapes, textures, and hand manipulation, improving alignment with visual_form (and tactile)."
    },
    {
      "key": "L29.F35",
      "layer": 29,
      "feat_id": 35,
      "label": "Body parts",
      "tldr": "Body parts",
      "dimensions": [
        "visual_form",
        "tactile"
      ],
      "primary_axis": "visual_form",
      "sae_r": 0.345,
      "llm_r": -0.128,
      "ceiling": 0.953,
      "max_act": 3710.9,
      "clamp_target": 7422.0,
      "n_active": 24,
      "neuronpedia": "https://www.neuronpedia.org/gemma-3-4b-it/29-gemmascope-2-res-16k/35",
      "top_nouns": [
        [
          "head",
          3710.9,
          "Sensorimotor/Head"
        ],
        [
          "arm",
          3286.6,
          "Sensorimotor/Hand_arm"
        ],
        [
          "forearm",
          3104.3,
          "Sensorimotor/Hand_arm"
        ],
        [
          "abdomen",
          2990.4,
          "Sensorimotor/Torso"
        ],
        [
          "eyes",
          2981.9,
          "Sensorimotor/Head"
        ],
        [
          "tummy",
          2867.8,
          "Sensorimotor/Torso"
        ],
        [
          "stomach",
          2672.8,
          "Sensorimotor/Torso"
        ],
        [
          "lung",
          2669.7,
          "Sensorimotor/Torso"
        ],
        [
          "hairline",
          2507.4,
          "Sensorimotor/Head"
        ],
        [
          "agility",
          2097.0,
          "Sensorimotor/Foot_leg"
        ]
      ],
      "corr": [
        [
          "visual_form",
          0.345,
          -0.128
        ],
        [
          "tactile",
          0.328,
          -0.088
        ],
        [
          "time",
          -0.31,
          0.366
        ],
        [
          "taste_smell",
          0.232,
          0.065
        ],
        [
          "thought",
          -0.226,
          0.634
        ],
        [
          "moral",
          -0.209,
          0.574
        ],
        [
          "color",
          0.2,
          0.033
        ],
        [
          "quantity",
          -0.185,
          0.403
        ],
        [
          "space",
          -0.179,
          0.333
        ],
        [
          "auditory",
          0.173,
          -0.007
        ],
        [
          "social",
          -0.081,
          0.499
        ],
        [
          "motion_self",
          0.05,
          0.418
        ],
        [
          "emotion",
          0.014,
          0.439
        ],
        [
          "polarity",
          0.002,
          0.512
        ]
      ],
      "interpretation": "Body-part detector: head, arm, forearm, abdomen, eyes, tummy, stomach, lung. A clean embodied-anatomy feature. Strongest on visual_form (+0.35) and tactile (+0.33). Note its Troche motion_self correlation is near zero: it fires on body parts, not on self-motion (the self-motion feature is L29.F132).",
      "steering_verdict": "Steers generations toward body parts and anatomy, improving alignment with the sensorimotor (visual_form, tactile) axes."
    }
  ]
}