Tomography — interpreted attractor axes across 12 language models

A catalog of 3,529 MLP-conductance attractor axes (7,058 signed poles) found by surrogate eigenflow censuses in the residual streams of GPT-2 small/medium/large/XL, Pythia 160M/410M/1.4B/2.8B/6.9B, Qwen3-0.6B/8B, and Qwen3.6-27B — each pole carrying an exemplar-derived natural-language label, a detection-validation score against a shuffled control, a kind categorization, and a cross-model family assignment. This repo supersedes davidafrica/gpt2-ur-feature-atlas (whose GPT-2 raw axes and census/causal results are re-exported here unchanged, now with their labels).

Layout

path	contents
attractors/<model>/<metric>/block_<l>.safetensors	axes — [k, d_model] fp16 unit axis vectors (residual-stream coordinates at the MLP-input hook site), cluster-id order
attractors/<model>/<metric>/block_<l>.json	per-axis census metadata + both poles' interpretation (label, confidence, detection balanced accuracy, control BA, validated flag, kind, coarse category, family id, example snippets)
attractors/<gpt2-model>/raw/…	the superseded atlas's GPT-2 raw axes, unchanged tensors, joined with the raw interpretation run's labels (poles_interp)
families.json	cross-model families from held-out word-profile clustering (Spearman ρ ≥ 0.70, ±0.05 stability), with LLM names and kind categorization
word_profiles.safetensors + word_types.json	[n_axes, 1533] held-slice word-type profiles (axis_gid order) and the shared type table
interp/	validation summary, kinds, protocol constants, raw-kind taxonomy
census_prior/	the superseded repo's census and causal-validation trees

<metric> is whitened (ε=0.1 shrinkage matched-filter surrogate eigenflow census) or raw (raw-projection census). Signs are arbitrary: each axis has a pos and a neg pole, interpreted separately.

Loading

import json
from safetensors.torch import load_file

d = load_file("attractors/pythia-160m/whitened/block_6.safetensors")
axes = d["axes"].float()                      # [k, 768], unit rows
meta = json.load(open("attractors/pythia-160m/whitened/block_6.json"))
for ax in meta["axes"]:
    print(ax["cluster_id"], ax["poles"].get("pos", {}).get("label"))

Projections follow the census convention: a = (h − μ_fit) · v with h the residual state at the MLP-input hook site (fp16-rounded) and μ_fit the model's fit-slice mean (pile-10k rows 0–101; see interp/protocol.py).

Domain-dense and rare-tail axes (sparse/)

333 attractor axes (666 signed poles) that the Pile censuses above never found, from a follow-up run that changed only the measure: the same frozen census on three domain-dense corpora (Python code, arXiv maths, UltraChat dialogue) and, on the Pile itself, a rare-event variant of the same flow that conditions on each direction's top 0.1% of states instead of its top 10%. Models: GPT-2 small and Qwen3-0.6B, four depth-spanning blocks each, with random-init floors under the identical protocol. An axis counts as novel when its largest absolute cosine against every axis in that model's Pile library (raw and whitened censuses, all blocks) is below 0.9.

Poles carry the same interpretation pipeline as the rest of this repo (exemplar → label → detection against a deranged-label control): 224/666 validated, categorised as domain structural 91, generic syntactic 100, semantic topical 33.

path	contents
sparse/poles.parquet	one row per signed pole: model, corpus, census metric, block, cluster, flow share, held-out lift, best cosine to the Pile library, label, detection balanced accuracy, validated flag, category
sparse/axes/<model>/<corpus>/block_<l>.safetensors	axes — [k, d_model] fp16 unit axis vectors in the same residual-stream coordinates as attractors/, and mu — the fit-slice mean of that corpus and block
sparse/axes/<model>/<corpus>/block_<l>.json	per-axis census metadata and both poles' interpretation
sparse/results.json	the pre-registered verdict statistics for both arms
sparse/protocol_sparse.py	the frozen constants of the follow-up run

<corpus> is code, math, chat or pile (the rare-tail arm). Axis projections use that corpus's own mu, not the Pile mean.

Diversity-maximising search directions (seq/)

1,193 directions that a diversity-maximising search finds above the same held-out acceptance floor the censuses in this repo use, but that the censuses themselves never returned. The search is a sequential orthogonalised ascent: repeatedly maximise the held-out gate lift subject to being orthogonal to everything already accepted, to a budget of 8x the censused count of that cell (capped at 64, floor 8). Models: GPT-2 small and Qwen3-0.6B, three depth-spanning blocks each, on Python code, arXiv maths, UltraChat dialogue and the Pile, with random-init floors run under the identical protocol and budget.

Pooled over the 22 cells that pass the floor-drift check, the search clears the floor 1,166 times against 280 censused axes (ratio 4.16), while the matched random-init arm produces 272 excess directions against the trained arm's 886. A direction counts as new here when its largest absolute cosine against every axis in that model's known set (its full Pile library plus the sparse/ axes) is below 0.9.

400 of them (a stratified subset, 800 signed poles) carry the same interpretation pipeline as the rest of this repo (exemplar to label to detection against a deranged-label control): 176/800 validated, categorised as domain structural 58, generic syntactic 86, semantic topical 32.

path	contents
seq/poles.parquet	one row per signed pole: model, corpus, frame, block, search slot, held-out lift, best cosine to the known set, whether it was labelled, label, detection balanced accuracy, validated flag, category
seq/axes/<model>/<corpus>/block_<l>.safetensors	axes — [k, d_model] fp16 unit direction vectors in the same residual-stream coordinates as attractors/, and mu — the fit-slice mean of that corpus and block
seq/axes/<model>/<corpus>/block_<l>.json	per-direction search metadata and both poles' interpretation
seq/counts.csv	per cell and frame: budget K, censused count, above-floor count, new count, for the trained and random-init arms
seq/results.json	the pre-registered Q1 count verdicts and the Q3 impact statistics
seq/validation.json	the Q2 labelling, detection and categorisation summary
seq/protocol_seq.py	the frozen constants of this run

<corpus> is code, math, chat or pile; frame is raw or whit (whitened). Projections use that corpus and block's own mu, not the Pile mean. Counts on random-init models are the floor, not a finding: the search clears its own floor on untrained weights too, which is why every count here is reported against its paired random-init cell.

Provenance

• Whitened censuses: exp_01m0tedzxhejkrq70x0pvcsay1 (GPT-2), exp_01m0xrq2h1fdertpjmkd9x80cs (Pythia/Qwen); raw scaling census: exp_01m0sy4fjhefv9tv0nwz9cx1nd; raw GPT-2 census + causal validation: exp_01m0rjapkveyqr5fytnhtwdzg8 / exp_01m0rjj4e8fdraqcf9vtcs2hnv.
• Interpretation (labels/validation/families/kinds): this repo's interp/ (labels: anthropic/claude-sonnet-5 via OpenRouter; detection validation against a within-model shuffled-control bar; corpus NeelNanda/pile-10k).
• Repo id: davidafrica/tomography.

A validated label certifies that it separates the pole's activating positions from non-activating ones (BA above the pooled shuffled-control 95th percentile); within-block specificity is not tested.

Training ladder (ladder/)

The same frozen census run on published Pythia training checkpoints: pythia-160m at 21 revisions (step0 … step143000, all 12 blocks) and pythia-410m at 8 revisions (blocks 4, 10, 16, 20). Per ladder/<model>/step<N>/<metric>/block_<l>.{safetensors,json}: the attractor axes of that checkpoint with census metadata and best_cos_to_final, the |cos| to the nearest final-checkpoint attractor. ladder/<model>/counts.json and nll_by_class.json give per-cell counts and held-slice NLL by token class; ladder/results.json is the joined formation/persistence analysis. Headline: whitened counts collapse to zero at steps 32–128, re-form by step 512–1000 and stay near their final level, while axis directions converge to their final positions only late (median best |cos| to the final library 0.28 at step 1000, 0.72 at step 14000, 0.98 at step 128000 for pythia-160m). See ladder/README.md for loading. Provenance: exp_01m1gsmvp5ejesbqkb43179j80.