soar-eleuther-i6-hierarchy/experiment_0-stats

experiment_0 — cached statistics (SOAR I-6)

Stage-01 tensor caches for soar-eleuther-i6-hierarchy/experiment_0, the "Implement Metrics" experiment of the SOAR I-6 project (hierarchy in SAEs).

Each file is one pass of NeelNanda/pile-10k (400 docs, context 128) through google/gemma-2-2b and its Matryoshka SAE on the residual stream, with every statistic the five hierarchy metrics need accumulated in a single sweep. Downloading these lets you skip cache_stats.py — the only GPU-heavy step — and run the whole analysis on CPU.

Files

File	Layer	SAE	Size
layer_03/exp0_stats.pt	3	gemma-2-2b/3-res-matryoshka-dc	~700 MB
layer_06/exp0_stats.pt	6	gemma-2-2b/6-res-matryoshka-dc	~700 MB
layer_12/exp0_stats.pt	12	gemma-2-2b/12-res-matryoshka-dc	~700 MB
layer_18/exp0_stats.pt	18	gemma-2-2b/18-res-matryoshka-dc	~700 MB
layer_24/exp0_stats.pt	24	gemma-2-2b/24-res-matryoshka-dc	~700 MB

Usage

git clone https://github.com/soar-eleuther-i6-hierarchy/experiment_0
cd experiment_0
pip install huggingface_hub torch plotly numpy

# one layer (~700 MB), straight into the path the scripts expect
hf download soar-eleuther-i6-hierarchy/experiment_0-stats \
    --repo-type dataset --include "layer_06/*" --local-dir outputs/

EXP0_LAYER=6 python3 run_metrics.py   # no model, no GPU
EXP0_LAYER=6 python3 visualize.py

Omit --include to pull all five layers (~3.4 GB).

What's inside a file

A torch.save dict (load with weights_only=False). The SAE's 32768 features are partitioned into 5 nested Matryoshka blocks by prefix length [128, 512, 2048, 8192, 32768] → B0=[0,128) B1=[128,512) B2=[512,2048) B3=[2048,8192) B4=[8192,32768). Pair-keyed entries are dicts keyed "0->1", "1->2", … for adjacent block pairs; feature indices inside a matrix are block-local.

Key	Shape / type	Meaning
schema_version	int (2)	v2 = BOS token excluded, plus the second-pass extras
fire_count	[32768]	tokens on which each feature fires (act > 1e-3)
total_tokens	int	tokens in the sweep, BOS excluded
token_counts, buckets	tensors	token-frequency counts and the 3 frequency buckets (0 = top 50% of mass, 1 = next 40%, 2 = rest)
pairs	list	the adjacent block pairs present
cofire	{pair: [P, C]}	tokens where parent and child both fire
cofire_by_bucket	{pair: [3, P, C]}	same, split by frequency bucket (metric 5)
g_parent_sum	{pair: [P, C]}	Σ ablation gain g_p = 2·a_p·⟨d_p, x−x̂⟩ + a_p²‖d_p‖² over the child's tokens (metric 2)
g_child_sum, err_sum_c	{block: [C]}	the child's own gain and base reconstruction error, the denominators for metric 2
fire_c_by_bucket	{block: [3, C]}	child fire counts per frequency bucket
within_cofire	{block: [C, C]}	within-block co-firing for sibling redundancy (metric 3); blocks 1–3 only — B4's 24576² does not fit
energy_cofire, union_count, union_energy, energy_total	{pair: ...}	second-pass (run_second_pass.py) accumulators
config	dict	layer, SAE id, block ranges, fire_threshold, n_docs, context size, thresholds — the full provenance of the run

Note: on Apple MPS the accumulators are float32 rather than float64 (MPS has no float64). Counts stay inside float32's exact-integer range and the reconstruction sums are only ever read as ratios, so this does not affect results — but don't assume float64.

Caveats

• B3→B4 is excluded from the pair accumulators for memory reasons (6144 × 24576).
• Sibling redundancy is unavailable for the B4 child, same reason.

Links

• Code: https://github.com/soar-eleuther-i6-hierarchy/experiment_0
• Dashboards: https://soar-eleuther-i6-hierarchy.github.io/experiment_0/