Hugging Face's logo

Datasets:
BurnyCoder
/
ouro-1.4b-neuron-atlas

Modalities:
Text
Formats:
json
Size:
< 1K
Tags:
interpretability
mechanistic-interpretability
looped-transformer
neurons
steering
ouro
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
Dataset Viewer
Auto-converted to Parquet Duplicate
pool
list
steered
list
candidates
list
rand_mean
list
u
dict
validation
list
[{"layer":10,"neuron":4225,"pole":0,"coherence":0.4733266830444336,"peak_share":0.1,"max_act":0.1269(...TRUNCATED)
[{"layer":23,"neuron":2726,"pole":1,"coherence":0.8179962038993835,"peak_share":0.25,"max_act":1.523(...TRUNCATED)
[{"layer":19,"neuron":3901,"pole":1,"coherence":0.6425589323043823,"peak_share":0.35,"max_act":2.640(...TRUNCATED)
[-0.018652206286787987,0.006196931935846806,0.009043669328093529,-0.009737014770507812,0.01351363211(...TRUNCATED)
{"19_3901_1":[-0.134128138422966,-0.07420429587364197,-0.04713372141122818,-0.025750156491994858,-0.(...TRUNCATED)
[{"key":"19_3901_1","alpha":4.0,"mean_dC":0.33672705642615225,"dz":2.6072523107255523,"dNLL_median":(...TRUNCATED)

Ouro-1.4B neuron atlas and steering results

Neuron-level atlas of ByteDance/Ouro-1.4B (revision 574fa66), a looped LM that applies its 24 layers 4 times. Produced by https://github.com/BurnyCoder/looped-transformers-llms-neuron-mechanistic-interpretability (see its README and reports/).

Files:

  • atlas.pt: torch dict for all 135,168 SwiGLU neurons (24 layers x 5632) and both poles (index 0 = positive, 1 = negative).
    • top_v/top_seq/top_pos/top_loop [24, 2, 5632, 40]: top-40 contexts, maximum over positions 4-255 x loops (the first 4 positions of every chunk are skipped).
    • lt_v/lt_seq/lt_pos [24, 4, 2, 5632, 20]: top-20 contexts per loop.
    • Per-loop mean, mean_abs, std, frac_pos [24, 4, 5632] (every 8th batch, 987,840 tokens per layer and loop).
    • exceed counts and the dead flag.
  • tokens_atlas.npy: uint16 [31250, 256], the token chunks that *_seq/*_pos index (Ouro tokenizer; documents drawn 90% from FineWeb-Edu sample-10BT shard 000 and 10% from OpenWebMath shard 0).
  • featured_neurons.jsonl: decoded top contexts (peak token marked <<>>) for the screened and validated candidates.
  • results/*.json: pre-registered confirmatory test, follow-ups, audit, replication and LLM-judge results.

Main result: 5/5 pre-registered neurons steer generation toward their concept, significantly against unsteered generation and against 100 random same-layer neurons clamped the same way (Holm-corrected), and the same neuron indices also steer Ouro-1.4B-Thinking (replication with a descriptive 20-neuron null).

Licenses and attribution: the tables and results produced by this project are MIT-licensed. tokens_atlas.npy and the decoded contexts in featured_neurons.jsonl derive from FineWeb-Edu (https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu) and OpenWebMath (https://huggingface.co/datasets/open-web-math/open-web-math), both ODC-By 1.0 and the Common Crawl Terms of Use (https://commoncrawl.org/terms-of-use), and remain subject to those terms. Activations come from Ouro-1.4B (Apache-2.0).
Downloads last month
77
Number of rows:
1
Total file size:
535 MB

Collection including BurnyCoder/ouro-1.4b-neuron-atlas