BrainAlign/brain-lm-alignment-ds001894

Brain–language-model alignment: ds001894 (whole-brain)

Lytle et al. 2019 — longitudinal word-level phonological processing in children scanned twice, at roughly 10 and 12 years old.

• Paper: https://www.nature.com/articles/s41597-019-0338-5
• Data: https://openneuro.org/datasets/ds001894/versions/1.4.2
• Generated: 2026-09-28
• Pipeline: https://github.com/suchirsalhan/cdl-representations-brains-babylms

Read this first: does the measurement work?

Every alignment number in this dataset is only as meaningful as the brain RDMs it was computed against. So before any model result, the same pipeline is asked whether anything stimulus-driven correlates with those RDMs — stimulus duration, intensity, word length, frequency, phoneme and syllable counts, an acoustic model of the audio where the stimuli are audio, and the study's own condition contrast — each tested by a permutation test that shuffles stimulus identity.

GATE: FAILED. 0/16 stimulus tests are significant after Holm correction — not the acoustic model of the audio the children actually heard, not the study's own experimental contrast.

The alignment numbers below are therefore uninterpretable as evidence about language models. They measure a representational geometry that does not demonstrably encode the stimuli. They are published for completeness and for whoever fixes the estimator, not as a result. Do not cite them as evidence that models fail to align with the developing brain.

Measured cause, from control/:

• RDM effective rank: None of None stimuli
• voxels per pattern: 120,865
• leading component vs the pattern's global signal: |ρ| = 0.48

Note that this is NOT ds003604's failure mode. There, the RDM effective rank was ~3 of 40-48 stimuli -- near-degenerate betas that could not express stimulus-level structure at all. The rank recorded above is a large fraction of the stimulus count, so these RDMs do carry stimulus structure and the control failing here means the specific controls tested did not reach significance, not that the measurement is uninterpretable. Check control/ for which controls ran: an acoustic or visual control needs the dataset's stimulus files present, and reports zero features if they are not.

What was built

60 task × session cells, each an RDM over the stimuli shared by that cell's subjects, with voxel patterns z-scored within run before aggregation (without that, the RDM measures scanner drift rather than language) and an inter-subject noise ceiling.

task	session	n_stim	ceiling_lower	ceiling_upper	ceiling_n
Phon	ses-11+	96	0.238099	0.424251	11
Phon	ses-11	96	0.244119	0.429977	11
Phon	ses-7	96	nan	nan	nan
Phon	ses-9	96	0.165375	0.503867	5
Orth	ses-7	96	nan	nan	nan
Phon	ses-7	96	nan	nan	nan
Orth	ses-11+	96	0.174681	0.550865	4
Orth	ses-11	96	0.237733	0.541529	5
Orth	ses-7	96	nan	nan	nan
Orth	ses-9	96	0.148507	0.613676	3
Phon	ses-11+	96	0.174681	0.550865	4
Phon	ses-11	96	0.237733	0.541529	5
Phon	ses-7	96	nan	nan	nan
Phon	ses-9	96	0.148507	0.613676	3
Orth	ses-11+	96	0.256457	0.431237	11
Orth	ses-11	96	0.254224	0.428718	11
Orth	ses-7	96	nan	nan	nan
Orth	ses-9	96	0.149473	0.488537	5
Phon	ses-11+	96	0.256457	0.431237	11
Phon	ses-11	96	0.254224	0.428718	11
Phon	ses-7	96	nan	nan	nan
Phon	ses-9	96	0.149473	0.488537	5
Orth	ses-7	96	nan	nan	nan
Phon	ses-7	96	nan	nan	nan
Orth	ses-11+	96	0.183085	0.551393	4
Orth	ses-11	96	0.237401	0.528923	5
Orth	ses-7	96	nan	nan	nan
Orth	ses-9	96	0.136605	0.600633	3
Phon	ses-11+	96	0.183085	0.551393	4
Phon	ses-11	96	0.237401	0.528923	5
Phon	ses-7	96	nan	nan	nan
Phon	ses-9	96	0.136605	0.600633	3
Orth	ses-11+	96	0.197762	0.396859	11
Orth	ses-11	96	0.211784	0.403963	11
Orth	ses-7	96	nan	nan	nan
Orth	ses-9	96	0.125687	0.472278	5
Phon	ses-11+	96	0.197762	0.396859	11
Phon	ses-11	96	0.211784	0.403963	11
Phon	ses-7	96	nan	nan	nan
Phon	ses-9	96	0.125687	0.472278	5
Orth	ses-7	96	nan	nan	nan
Phon	ses-7	96	nan	nan	nan
Phon	ses-11+	96	0.116967	0.512649	4
Phon	ses-11	96	0.174403	0.502922	5
Phon	ses-7	96	nan	nan	nan
Phon	ses-9	96	0.126193	0.592164	3
Phon	ses-11+	96	0.253069	0.431488	11
Phon	ses-11	96	0.25297	0.433396	11
Phon	ses-7	96	nan	nan	nan
Phon	ses-9	96	0.158864	0.490917	5
Orth	ses-7	96	nan	nan	nan
Phon	ses-7	96	nan	nan	nan
Orth	ses-11+	96	0.174262	0.545734	4
Orth	ses-11	96	0.213686	0.52405	5
Orth	ses-7	96	nan	nan	nan
Orth	ses-9	96	0.165076	0.607179	3
Phon	ses-11+	96	0.174262	0.545734	4
Phon	ses-11	96	0.213686	0.52405	5
Phon	ses-7	96	nan	nan	nan
Phon	ses-9	96	0.165076	0.607179	3

Model grid: 15 families, 8397 alignment rows across 5 cells.

mean noise ceiling	0.191
best alignment anywhere	0.0431
as a fraction of ceiling	34.5%
families equivalent to zero (TOST ±0.05)	14/15
Pythia scale trend	ρ = +0.013, p = 0.96

Per family

family	n_checkpoints	rsa_mean	rsa_sd	rsa_abs_max	frac_of_ceiling_abs_max	p_equivalence_tost
pythia-160m-full	21	0.0091	0.0098	0.0405	0.2924	0.0018
pico-decoder-tiny	21	0.0071	0.0089	0.0403	0.3449	0.0012
pythia-1.4b-full	21	0.0065	0	0.0253	nan	nan
pythia-1b-full	1	0.0052	0.0032	0.0089	0.0509	0
pythia-410m-full	21	0.0041	0.0093	0.0325	0.2287	0.0011
pico-decoder-large	21	0.003	0.012	0.0374	0.2978	0.0021
pythia-70m-full	21	0.0026	0.0093	0.0339	0.2697	0.001
pico-decoder-small	21	0.0025	0.0112	0.0359	0.2857	0.0017
pico-decoder-medium	21	0.0022	0.0123	0.0364	0.2812	0.0021
beetle-fineweb3-eng	19	-0.001	0.0034	0.0298	0.2546	0
beetle-humanscale-eng	18	-0.0026	0.0054	0.0347	0.2758	0.0002
babylm-gpt2	9	-0.0063	0.005	0.0235	0.2012	0.0002
babylm-gpt2-3	9	-0.0244	0.0128	0.0413	0.3289	0.0124
babylm-gpt2-5	9	-0.025	0.0127	0.0418	0.3325	0.0131
babylm-gpt2-7	9	-0.0259	0.013	0.0431	0.3431	0.0154

Dataset-specific notes

The only longitudinal dataset here: the same children at two timepoints (ses-T1, ses-T2), which is the closest real analogue to a language model's checkpoint trajectory. Per-subject age at scan is available. Trial types cross orthographic with phonological similarity (O+P+/O+P-/O-P+/O-P-), so Phon and Orth contrasts are decorrelated by design. ses-T2 has only the VV tasks.

Files

path	what	present here
alignment_by_checkpoint.csv	every model × checkpoint × cell, with ceiling	✓
alignment_by_family.csv	per family, with equivalence tests	✓
alignment_by_cell.csv	per task × session	✓
ceilings_ds001894.csv	noise ceiling per cell	✓
control/	the positive control and RDM dimensionality — the gate	✓
scale_ladder.csv	the Pythia 70M→1.4B scale test	✓
fig_*.pdf, fig_*.png	figures	✓

Method

Representational similarity analysis. For each cell, a brain RDM over stimuli (correlation distance between per-stimulus GLM beta patterns, within-run z-scored, aggregated across subjects) is compared by Spearman correlation with a model RDM over the same stimuli, taken from each checkpoint's hidden states. Alignment is reported raw and as a fraction of the inter-subject noise ceiling, and judged against a null built from the PARC suite — 18 models differing only by random seed, which is what 'no effect' looks like on this measurement.

Null and fixation trials are excluded from the stimulus set. For paired designs the stimulus identity is the pair, not either word alone.