Dataset Viewer

The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.

Smol Neuron Studio — unit catalog

Catalog behind Smol Neuron Studio: for every MLP neuron ("unit") of the model, the text where it fires hardest, activation statistics, the tokens it promotes and suppresses (logit lens) and, for part of them, a short human-readable label.

One folder per model. Each holds a single SQLite file, catalog.sqlite.

Folder Model Units Labeled Text read
SmolLM2-135M-Instruct HuggingFaceTB/SmolLM2-135M-Instruct 46,080 759 1,280,000 tokens

How it was built

  • Units: the input of mlp.down_proj in every decoder layer, read with PyTorch hooks.
  • Text: fragments of 64 tokens from FineWeb-Edu (HuggingFaceFW/fineweb-edu/sample/10BT/000_00000.parquet). Token 0 of each fragment (the attention sink) is skipped.
  • Per unit: top / bottom examples, mean, |mean|, std, max, min, fraction of tokens active, and logit lens (down_proj column → final norm → unembedding).
  • Labels: written by Claude from the top examples and logit lens, with a confidence score; they are hypotheses, not ground truth. unclear marks units without a readable pattern.

Tables

meta, fragments, unit_stats, unit_examples, unit_logits, unit_labels.

License

Text snippets come from FineWeb-Edu (ODC-By 1.0); this catalog keeps the same license.

Downloads last month
44

Space using lorcapoul/smol-neuron-studio-catalog 1