Neuronpedia MLP Neuron Dashboards

Activation dashboards for every MLP neuron of three open models, computed on the same 100M-token corpus. This is a copy of the files that Neuronpedia serves from https://dashboards.neuronpedia.org/v2/<model>/. The files are byte-identical.

Model	Hugging Face id	Layers	Neurons per layer	Neurons
qwen3-32b	Qwen/Qwen3-32B	64	25,600	1,638,400
gemma-4-31b-it	google/gemma-4-31B-it	60	21,504	1,290,240
llama3.3-70b-it	meta-llama/Llama-3.3-70B-Instruct	80	28,672	2,293,760

A neuron's activation is act = act_fn(gate) * up, the input to down_proj. "Gate open" means gate > 0. Position 0 of each document is never counted.

Layout

v2/
  corpus/
    docs.parquet          one row per document; examples point at a row by "doc"
    docs.txt              the same texts as UTF-8, concatenated, for byte-range reads
    docs_offsets.npy      document d is docs.txt bytes [offsets[d], offsets[d+1]), int64
    stats.json            documents and tokens per corpus slice
    removed_docs.json     ids of the removed code documents (see "Corpus and licenses")
    code_sources.json     license, repo and path of each code document that is kept
  <model>/
    meta.json             run settings and the record format ("format" key)
    neuron_percentiles.safetensors   per-neuron percentiles of |act|, for ablation thresholds
    corpus/
      tokens.npy          all corpus tokens for this model's tokenizer, concatenated
      offsets.npy         document d is tokens[offsets[d]:offsets[d+1]]
      vocab.json          token id -> string
    neurons/
      layer_XX.json.gz    one gzip member per neuron, in neuron order
      layer_XX.idx        N + 1 byte offsets into the .json.gz, little-endian uint64
      layer_XX.index.json the same offsets as JSON
    raw/
      layer_XX.npz        the arrays the records were built from

Read one neuron

Each gzip member holds one neuron, so you need only its byte range:

import gzip, json
import numpy as np

model, layer, neuron = "qwen3-32b", 31, 1234
base = f"v2/{model}/neurons/layer_{layer:02d}"
offsets = np.fromfile(f"{base}.idx", dtype="<u8")
with open(f"{base}.json.gz", "rb") as f:
    f.seek(int(offsets[neuron]))
    member = f.read(int(offsets[neuron + 1] - offsets[neuron]))
record = json.loads(gzip.decompress(member))
print(record["stats"], record["logits"]["pos"][:5])

To download one layer only:

hf download neuronpedia/dashboards --repo-type dataset --include "v2/qwen3-32b/neurons/layer_31.*" --local-dir .

Record format

• stats: max, min, mean, std, sigma_calib, frac_gate_open, echo_abs_share, n_hist_tokens.
• quantiles: activation values v at quantiles q.
• hist: counts on every 8th token, split into gate open and gate closed, positive and negative, 66 bins per sign. Bin 0 is |act| / sigma < 1e-3, bins 1 to 64 are log-spaced up to 1e3 * sigma, and bin 65 is above that.
• Example lists:
  • top and bottom: gate-open examples with the largest and most negative act, one per document.
  • echo_top and echo_bottom: the same for gate-closed examples.
  • bands: random examples with act in [lo, hi), where the band edges are the calibration quantiles q_lo and q_hi.
• Each example has:
  • doc (row in docs.parquet) and pos (the peak token position in the document);
  • act, gate and up at the peak token;
  • start (the document position of the first window token);
  • tokens (model token ids) and acts (activation at each window token);
  • gate_closed (one character per window token, where 1 means gate <= 0).
• logits: [token id, token, value] for the neuron's direct effect on the unembedding (W_U @ (final_norm * post_ffn_norm * down_proj[:, n]), with the norms applied as scales only).

meta.json repeats this under its format key.

Corpus and licenses

The corpus has 191,484 documents:

Slice	Share of tokens	Source	License
English web text	70%	FineWeb (sample-10BT)	ODC-By 1.0
Ten other languages	15%	FineWeb-2	ODC-By 1.0
Code	10%	codeparrot/github-code-clean	license of each file's repo (see below)
Math	5%	FineMath (finemath-4plus)	ODC-By 1.0

FineWeb, FineWeb-2 and FineMath are released under the Open Data Commons Attribution License (ODC-By) v1.0. Their use is also subject to Common Crawl's Terms of Use. The web pages themselves stay the property of their owners.

Code. Each code document is the start of one GitHub file. We kept only files whose repository license is permissive: MIT, Apache-2.0, BSD-2-Clause, BSD-3-Clause, ISC, Unlicense, CC0-1.0 or Artistic-2.0. The other 7,621 code documents (GPL, LGPL, AGPL, MPL and EPL) were removed:

• their text in docs.parquet and their tokens in <model>/corpus/tokens.npy are blank;
• examples on them are dropped from the neuron records, so some lists have fewer examples;
• raw/ marks their entries with doc = -1.

Document ids did not change. The stats, quantiles and histograms still count the tokens of the removed documents, because they hold no text. corpus/code_sources.json gives the license, repo and path of each code document that is kept. Keep that notice when you reuse the code text.

To have a document removed, open a discussion on this repository.