by-model/pico-decoder-tiny/README.md

pico-decoder-tiny

pico decoder, smallest of the scale ladder — pico-lm/pico-decoder-tiny · dense commit-history checkpoints.

21 checkpoints, steps 0–126000.

⚠️ The brain alignment numbers on this page are confounded

brain_alignment.csv, every brain_* column in checkpoints.csv, and ablation_alignment.csv are not interpretable as a result about this model.

In ds003604 each stimulus appears in exactly one scanner run, so run membership is perfectly confounded with stimulus identity. "Different run" predicts brain dissimilarity at Spearman +0.49 to +0.87 across all 12 task × session cells, while no stimulus property predicts it at all (trial type ≈ −0.02, length ≈ 0, lexical overlap ≈ 0). A language model cannot represent a scanner run, so its RSA against that structure is ≈0 by construction.

Ruled out as explanations: cohort size (40 vs 98 subjects agree at rho 0.928) and layer choice (flat across all 12–14 layers). Within-run normalisation removes the confound (+0.562 → −0.041) and leaves alignment at +0.022, not significant.

interp_*, localisation_*, behaviour and ablation_behaviour on this page are unaffected — they are computed from the model alone and never touch the fMRI data. See the repo README for the full account.

Headline numbers

metric	value	safe to use?
brain RSA (mean over checkpoints × 12 cells)	-0.0167	❌ confounded
brain RSA vs training step	rho -0.016, p 0.799 (n=252)	❌ confounded
interp: participation ratio (effective dim / hidden)	0.2209	✅
interp: gini (activation sparsity)	0.2042	✅
interp: hoyer sparsity	0.0933	✅
interp: CKA to previous checkpoint	0.9353	✅
localisation: selectivity index	0.5459	✅
localisation: mean overlap with other phenomena	0.0030	✅
localisation: active layers	7.0595	✅
localisation: layer centre of mass	0.6129	✅
behaviour: minimal-pair accuracy (chance 0.5)	0.6311	✅

What this model shows

• Minimal-pair accuracy 0.631, modestly above chance.
• Participation ratio 0.221: the representation uses ~22% of its available dimensions. Mid-range for this release.
• Selectivity 0.546 — like every model in this release, phenomena are only moderately isolated. Nothing here shows sharp, dedicated linguistic circuitry.
• CKA to the previous checkpoint reaches 0.9911 by the end: the representation has stopped changing, so late checkpoints are near-duplicates.
• Brain alignment: no claim is made, see the warning above.

Files

file	contents	confounded?
behaviour.csv	minimal-pair accuracy per phenomenon per checkpoint	no
brain_alignment.csv	RSA per task × session × checkpoint	yes — whole file
checkpoints.csv	this model's rows of the main per-checkpoint table (all three axes)	brain_* columns only
interp_layerwise.csv	the same metrics per layer per checkpoint	no
interp_mechanistic.csv	per-checkpoint representation metrics	no
localisation_isolation.csv	per-phenomenon circuit localisation	no
localisation_onset.csv	localisation onset step per phenomenon	no