sardines-mlp

Trained runs for sardines-mlp: a reproduction of Elhage et al. 2022, Toy Models of Superposition. The toy model squeezes n sparse features into m < n dimensions and reads them back with the same matrix, x' = ReLU(WᵀWx + b). The code, the full write-up and the figures are on GitHub.

Each run folder holds one sweep, trained as a batch of independent instances:

• model.safetensors: W (instances × m × n) and b (instances × n).
• sweep.safetensors: every instance's feature probability 1 − S and importances.
• metrics.json: the full training record (config, seeds, coordinates, evaluation history, baselines, code state, environment).
• training.png: a diagnostic plot.
• analysis/: per-instance and per-feature geometry tables, a diagnostic plot and a summary.

Sweep	n → m	Instances	What it shows
intro-{relu,linear}-seed{0..4}	20 → 5	7 sparsities	The ReLU model keeps up to 13–14 features in 5 dimensions; the linear model keeps 5
polygons-{relu,linear}-seed{0..4}	5 → 2	21 densities	A regular pentagon in all five seeds once 1 − S ≤ 0.05
phase-{relu,linear}-seed{0..4}	2 → 1	24 × 24 grid	When the second feature is dropped, gets the dimension, or forms an antipodal pair
dimensionality-{relu,linear}-seed{0..4}	400 → 30	40 sparsities	Half a dimension per feature over a range of sparsities; features at ½, ⅖ and ⅜

The compare-* folders hold each sweep's geometry across the five seeds, the numbers the README figures are drawn from.

To load a run, with the sardines-mlp package installed:

from pathlib import Path
from huggingface_hub import snapshot_download
from sardines_mlp import checkpoint

folder = Path(snapshot_download('coeiico/sardines-mlp', allow_patterns='polygons-relu-seed0/*')) / 'polygons-relu-seed0'
run = checkpoint.load(folder)  # run.model, run.sweep, run.record

These are evaluation weights, not resumable checkpoints. The intro, polygons and phase runs were trained on the CPU and the dimensionality runs on CUDA (an RTX 3060). A seed fixes a run on one kind of device.