candle-mi 0.1.5

Mechanistic interpretability for language models in Rust, built on candle
{
  "model_id": "google/gemma-2-2b",
  "prompt": "The capital of France is",
  "hook_layer": 0,
  "hook_name": "blocks.0.hook_resid_post",
  "sae_repo": "google/gemma-scope-2b-pt-res",
  "sae_npz_path": "layer_0/width_16k/average_l0_105/params.npz",
  "d_in": 2304,
  "d_sae": 16384,
  "architecture": "JumpReLU",
  "tokens": [
    "<bos>",
    "The",
    "?capital",
    "?of",
    "?France",
    "?is"
  ],
  "n_tokens": 6,
  "reconstruction_mse": 6.912656307220459,
  "n_active_last_pos": 14,
  "top_features_last_pos": [
    {
      "index": 10492,
      "value": 44.07236099243164
    },
    {
      "index": 8920,
      "value": 16.872859954833984
    },
    {
      "index": 6887,
      "value": 8.597606658935547
    },
    {
      "index": 5542,
      "value": 6.81044864654541
    },
    {
      "index": 6085,
      "value": 6.258038520812988
    },
    {
      "index": 6735,
      "value": 5.713671684265137
    },
    {
      "index": 14856,
      "value": 3.5655159950256348
    },
    {
      "index": 11517,
      "value": 3.366048812866211
    },
    {
      "index": 10034,
      "value": 3.119041681289673
    },
    {
      "index": 688,
      "value": 2.0419840812683105
    }
  ],
  "resid_last_first10": [
    -0.46773838996887207,
    0.018950626254081726,
    -0.4665502905845642,
    -0.07082527875900269,
    -0.1438910961151123,
    0.10148066282272339,
    -0.32458484172821045,
    -2.141545057296753,
    -1.4411273002624512,
    0.14355504512786865
  ],
  "encoded_last_first10": [
    -0.0,
    -0.0,
    -0.0,
    -0.0,
    -0.0,
    -0.0,
    -0.0,
    -0.0,
    -0.0,
    -0.0
  ],
  "decoded_last_first10": [
    -0.6751449108123779,
    -0.20853310823440552,
    -0.24843807518482208,
    -0.0810665637254715,
    -0.19063863158226013,
    0.3584142029285431,
    -0.3056459426879883,
    -2.5966145992279053,
    -1.1539592742919922,
    0.2002873420715332
  ],
  "resid_last_norm": 61.81248474121094,
  "encoded_last_norm": 49.715003967285156,
  "decoded_last_norm": 59.745872497558594
}