replicaverse/replica-predictors

Per-word reading predictors

Precomputed per-word predictors for computational models of human sentence processing: word surprisal, several kinds of entropy, and unigram surprisal, scored over reading corpora (eye tracking, self-paced reading, N400, maze).

Every row says what text unit it describes and what unit its number is in:

dataset	tag	predictor	stimulus_id	unit_index	text_unit	value	value_unit
provo	gpt2_ft_entropy_provo	token_entropy	provo_1	4	Apple	8.4213	bits

Counts below describe cells indexed in tags.parquet. Cells are unique tags. Rows sum n_rows over (tag, predictor, file), so one word can contribute to several predictor rows.

• 2,111 cells across 58 corpora, 142,164,743 scored rows
• 25 methods x 26 models x 158 predictor keys

Start here: tags.parquet

One row per (tag, predictor, file), with the model, method and configuration decoded out of the tag, so you can find a cell by what it is instead of by guessing at its name.

import pandas as pd
idx = pd.read_parquet("hf://datasets/replicaverse/replica-predictors/tags.parquet")

# Every word-level Renyi-1/2 entropy cell computed with 512 samples.
idx[(idx.method == "mc_word_entropy") & (idx.alpha == 0.5) & (idx.n_samples == 512)]

# What is available for one corpus?
idx[idx.dataset == "provo"][["model", "method", "predictor", "value_unit", "n_rows"]]

# Which file holds the raw sample arrays behind a scalar cell?
idx[idx.tag == "gpt2_mcword_entropy_N512_M20_a1_s0_provo"][["predictor", "file"]]

column	meaning
tag	the cell id, as it appears in the tag column of the data files
file	the file in this repo that holds these rows
dataset	corpus
model	the model or frequency table the cell was computed from
model_kind	lm or frequency_table
method	how the value was computed (see Methods)
predictor	the value key
value_unit	bits, nats, dimensionless, identifier or text
n_stimuli	stimuli covered
n_rows	rows, which is also the number of scored units
alpha, n_samples, max_tokens, beam_width, max_words, seed	the cell's configuration, null where the method has no such setting
provenance	where the columns below come from: manifest (the runner's sidecar, written for cells computed since 2026-09-06), unit_metadata (the Monte Carlo families stamp device, torch build and chunk size on every unit), records (the frozen compute environment, the job log that wrote the cell, the scheduler's GPU record, the pickle's write time and the model cache), or none
chunk_size, device, cuda, torch, transformers, python, replica	the realization: how many samples were drawn at once, the GPU the model ran on, and the package versions. A sample is only reproducible on the same device and torch build
code_commit, code_dirty, source_sha256	the commit of the source tree the cell was computed at, whether that tree had uncommitted changes, and a content hash of src/replica that identifies the code even without a commit
model_revision, computed_at	the Hugging Face revision of the model snapshot and when the cell was written

Cells published before 2026-09-08 predate the manifest. The environment columns are filled only for the Monte Carlo families, which carried their device and torch build on every unit. Their code_commit was established afterwards from the schedulers' job accounting, the checkouts' reflogs, and a file-by-file match of the compute tree against the repository; the 50 cells whose tree was overwritten before it could be matched stay null, as does every other pre-manifest cell. The history.parquet file records, for every cell, the commit that first published it and the commit that last changed it: the publishing commit is not the computing commit, but it is the one record every cell has.

Row schema

The per-corpus files are long format: one row per (tag, predictor, stimulus_id, unit_index).

column	type	meaning
dataset	string	corpus name, which is also the file name of the per-corpus tables
tag	string	cell id: model, method and configuration, e.g. gpt2_surprisal_provo
predictor	string	which quantity the row carries, e.g. token_entropy
stimulus_id	string	stimulus (passage or sentence) within the corpus
unit_index	int32	0-based position of the text unit within the stimulus
text_unit	string	the text unit itself: the word, or the character for the meco_char_* corpora
value	double	the number, NaN where undefined
value_unit	string	measurement unit of value

text_unit is the surface string the predictor was computed on, so a row can be read on its own and a join against the corpus is only needed to add reading times. It is null wherever that string cannot be established, which is a small and enumerated set of stimuli (see Data notes); a non-null text_unit is the string the value was computed on, never a guess.

Predictors and units

Log bases are mixed on purpose: log_p_observed is a signed natural log-probability, kept in the base it is computed in, and every entropy is in bits. value_unit is what disambiguates them, and the same information is attached to the array columns as Parquet field metadata.

All indexed predictor keys and units

predictor	value_unit	cells	rows
continuation_entropy	bits	170	3,211,755
continuation_entropy_buggy	bits	1	1,931
joint_word_entropy	bits	22	700,591
mc_word_entropy	bits	312	7,057,470
next_word_entropy	bits	126	1,448,928
next_word_renyi_entropy_alpha0.5	bits	101	1,430,124
token_entropy	bits	86	1,371,270
token_renyi_entropy_alpha0.5	bits	87	1,403,256
token_renyi_entropy_alpha2.0	bits	83	1,365,466
unigram_surprisal	bits	154	2,342,173
decontextualised_distance	dimensionless	58	1,231,728
expected_decontextualised_distance	dimensionless	58	1,231,728
expected_similarity_tokens	dimensionless	25	606,073
information_value_context_informativeness_1gram	dimensionless	54	639,428
information_value_context_informativeness_1gram_pos	dimensionless	54	639,428
information_value_context_informativeness_2gram	dimensionless	54	639,428
information_value_context_informativeness_2gram_pos	dimensionless	54	639,428
information_value_context_informativeness_3gram	dimensionless	54	639,428
information_value_context_informativeness_3gram_pos	dimensionless	54	639,428
information_value_context_informativeness_cosine	dimensionless	54	639,428
information_value_context_informativeness_euclidean	dimensionless	54	639,428
information_value_deviation_1gram	dimensionless	54	639,428
information_value_deviation_1gram_pos	dimensionless	54	639,428
information_value_deviation_2gram	dimensionless	54	639,428
information_value_deviation_2gram_pos	dimensionless	54	639,428
information_value_deviation_3gram	dimensionless	54	639,428
information_value_deviation_3gram_pos	dimensionless	54	639,428
information_value_deviation_cosine	dimensionless	54	639,428
information_value_deviation_euclidean	dimensionless	54	639,428
information_value_expected_1gram	dimensionless	54	639,428
information_value_expected_1gram_pos	dimensionless	54	639,428
information_value_expected_2gram	dimensionless	54	639,428
information_value_expected_2gram_pos	dimensionless	54	639,428
information_value_expected_3gram	dimensionless	54	639,428
information_value_expected_3gram_pos	dimensionless	54	639,428
information_value_expected_context_informativeness_1gram	dimensionless	54	639,428
information_value_expected_context_informativeness_1gram_pos	dimensionless	54	639,428
information_value_expected_context_informativeness_2gram	dimensionless	54	639,428
information_value_expected_context_informativeness_2gram_pos	dimensionless	54	639,428
information_value_expected_context_informativeness_3gram	dimensionless	54	639,428
information_value_expected_context_informativeness_3gram_pos	dimensionless	54	639,428
information_value_expected_context_informativeness_cosine	dimensionless	54	639,428
information_value_expected_context_informativeness_euclidean	dimensionless	54	639,428
information_value_expected_cosine	dimensionless	54	639,428
information_value_expected_euclidean	dimensionless	54	639,428
information_value_max_1gram	dimensionless	54	639,428
information_value_max_1gram_pos	dimensionless	54	639,428
information_value_max_2gram	dimensionless	54	639,428
information_value_max_2gram_pos	dimensionless	54	639,428
information_value_max_3gram	dimensionless	54	639,428
information_value_max_3gram_pos	dimensionless	54	639,428
information_value_max_cosine	dimensionless	54	639,428
information_value_max_euclidean	dimensionless	54	639,428
information_value_mean_1gram	dimensionless	54	639,428
information_value_mean_1gram_pos	dimensionless	54	639,428
information_value_mean_2gram	dimensionless	54	639,428
information_value_mean_2gram_pos	dimensionless	54	639,428
information_value_mean_3gram	dimensionless	54	639,428
information_value_mean_3gram_pos	dimensionless	54	639,428
information_value_mean_cosine	dimensionless	54	639,428
information_value_mean_euclidean	dimensionless	54	639,428
information_value_min_1gram	dimensionless	54	639,428
information_value_min_1gram_pos	dimensionless	54	639,428
information_value_min_2gram	dimensionless	54	639,428
information_value_min_2gram_pos	dimensionless	54	639,428
information_value_min_3gram	dimensionless	54	639,428
information_value_min_3gram_pos	dimensionless	54	639,428
information_value_min_cosine	dimensionless	54	639,428
information_value_min_euclidean	dimensionless	54	639,428
information_value_ooc_deviation_1gram	dimensionless	54	639,428
information_value_ooc_deviation_1gram_pos	dimensionless	54	639,428
information_value_ooc_deviation_2gram	dimensionless	54	639,428
information_value_ooc_deviation_2gram_pos	dimensionless	54	639,428
information_value_ooc_deviation_3gram	dimensionless	54	639,428
information_value_ooc_deviation_3gram_pos	dimensionless	54	639,428
information_value_ooc_deviation_cosine	dimensionless	54	639,428
information_value_ooc_deviation_euclidean	dimensionless	54	639,428
information_value_ooc_expected_1gram	dimensionless	54	639,428
information_value_ooc_expected_1gram_pos	dimensionless	54	639,428
information_value_ooc_expected_2gram	dimensionless	54	639,428
information_value_ooc_expected_2gram_pos	dimensionless	54	639,428
information_value_ooc_expected_3gram	dimensionless	54	639,428
information_value_ooc_expected_3gram_pos	dimensionless	54	639,428
information_value_ooc_expected_cosine	dimensionless	54	639,428
information_value_ooc_expected_euclidean	dimensionless	54	639,428
information_value_ooc_max_1gram	dimensionless	54	639,428
information_value_ooc_max_1gram_pos	dimensionless	54	639,428
information_value_ooc_max_2gram	dimensionless	54	639,428
information_value_ooc_max_2gram_pos	dimensionless	54	639,428
information_value_ooc_max_3gram	dimensionless	54	639,428
information_value_ooc_max_3gram_pos	dimensionless	54	639,428
information_value_ooc_max_cosine	dimensionless	54	639,428
information_value_ooc_max_euclidean	dimensionless	54	639,428
information_value_ooc_mean_1gram	dimensionless	54	639,428
information_value_ooc_mean_1gram_pos	dimensionless	54	639,428
information_value_ooc_mean_2gram	dimensionless	54	639,428
information_value_ooc_mean_2gram_pos	dimensionless	54	639,428
information_value_ooc_mean_3gram	dimensionless	54	639,428
information_value_ooc_mean_3gram_pos	dimensionless	54	639,428
information_value_ooc_mean_cosine	dimensionless	54	639,428
information_value_ooc_mean_euclidean	dimensionless	54	639,428
information_value_ooc_min_1gram	dimensionless	54	639,428
information_value_ooc_min_1gram_pos	dimensionless	54	639,428
information_value_ooc_min_2gram	dimensionless	54	639,428
information_value_ooc_min_2gram_pos	dimensionless	54	639,428
information_value_ooc_min_3gram	dimensionless	54	639,428
information_value_ooc_min_3gram_pos	dimensionless	54	639,428
information_value_ooc_min_cosine	dimensionless	54	639,428
information_value_ooc_min_euclidean	dimensionless	54	639,428
infovalue_layer0	dimensionless	41	665,808
infovalue_layer1	dimensionless	41	665,808
infovalue_layer10	dimensionless	41	665,808
infovalue_layer11	dimensionless	41	665,808
infovalue_layer12	dimensionless	41	665,808
infovalue_layer13	dimensionless	41	665,808
infovalue_layer14	dimensionless	41	665,808
infovalue_layer15	dimensionless	41	665,808
infovalue_layer16	dimensionless	41	665,808
infovalue_layer17	dimensionless	41	665,808
infovalue_layer18	dimensionless	41	665,808
infovalue_layer19	dimensionless	41	665,808
infovalue_layer2	dimensionless	41	665,808
infovalue_layer20	dimensionless	41	665,808
infovalue_layer21	dimensionless	41	665,808
infovalue_layer22	dimensionless	41	665,808
infovalue_layer23	dimensionless	41	665,808
infovalue_layer24	dimensionless	41	665,808
infovalue_layer3	dimensionless	41	665,808
infovalue_layer4	dimensionless	41	665,808
infovalue_layer5	dimensionless	41	665,808
infovalue_layer6	dimensionless	41	665,808
infovalue_layer7	dimensionless	41	665,808
infovalue_layer8	dimensionless	41	665,808
infovalue_layer9	dimensionless	41	665,808
next_word_information_value	dimensionless	25	606,073
next_word_information_value_first	dimensionless	25	606,073
next_word_information_value_orth	dimensionless	6	28,953
next_word_information_value_pos	dimensionless	6	28,953
log_p_layers	nats	204	6,492,037
log_p_observed	nats	735	13,695,105
log_p_samples_nats	nats	482	10,269,225
next_log_probs	nats	21	62,355
similarity_adjusted_surprisal	nats	25	606,073
similarity_adjusted_surprisal_first	nats	25	606,073
similarity_adjusted_surprisal_orth	nats	6	28,953
similarity_adjusted_surprisal_pos	nats	6	28,953
information_value_context_informativeness_length	tokens	54	639,428
information_value_deviation_length	tokens	54	639,428
information_value_expected_context_informativeness_length	tokens	54	639,428
information_value_expected_length	tokens	54	639,428
information_value_max_length	tokens	54	639,428
information_value_mean_length	tokens	54	639,428
information_value_min_length	tokens	54	639,428
information_value_ooc_deviation_length	tokens	54	639,428
information_value_ooc_expected_length	tokens	54	639,428
information_value_ooc_max_length	tokens	54	639,428
information_value_ooc_mean_length	tokens	54	639,428
information_value_ooc_min_length	tokens	54	639,428

value_unit	means
nats	log base e. log_p_observed is a log-probability, so it is <= 0; negate and divide by ln 2 for surprisal in bits
bits	log base 2, >= 0
dimensionless	a cosine distance or a probability-like ratio, no log base
identifier	not a measurement: a hash used to group samples
text	not a measurement: a surface string

Tag grammar

<model>_<method><config>_<dataset>
   |        |        |        |
   |        |        |        `- corpus, also the file name
   |        |        `- configuration tokens, method dependent
   |        `- how the value was computed
   `- HuggingFace model id with "/" written as "_", or a frequency table name

gpt2_mcword_entropy_N512_M20_a1_s0_provo is gpt2, word entropy by Monte Carlo, 512 samples, at most 20 subword steps per sampled word, Renyi order 1 (Shannon), seed 0, on provo. The pieces are already split out in tags.parquet, so nothing needs to parse this string.

token	column	meaning
N<n>	n_samples	Monte-Carlo samples drawn per unit
L<n>	max_tokens	continuation length, in subword tokens
M<n>	max_tokens	cap on subword steps taken while extending one sampled word
K<n>	beam_width / max_words	beam width for charbeam, look-ahead depth in words for fwd_word_lookahead
a<x>	alpha	Renyi order. a1 is the Shannon limit
s<n>	seed	random seed

Some fwd_word_lookahead_entropy_samples/ files end in .shard<start>_<end>: one slice of a corpus that was scored in pieces, covering stimuli start to end. The part before it is the cell id, and tags.parquet groups the slices under that id while its file column keeps them apart.

Methods

method	predictors it carries	settings used	cells	corpora	models
char_beam	log_p_observed	beam_width=5	43	28	2
continuation_entropy	continuation_entropy, log_p_samples_nats	n_samples=64, 512; max_tokens=5, 10, 15; seed=0	170	57	2
continuation_entropy_buggy	continuation_entropy_buggy	n_samples=512; max_tokens=5; seed=0	1	1	1
first_token_entropy	next_log_probs, token_entropy	none	86	58	2
forward_word_lookahead_entropy	joint_word_entropy	n_samples=512; max_tokens=30; max_words=5; seed=0	22	18	2
information_value	108 keys; see predictor table or tags.parquet	n_samples=100; max_tokens=60; seed=0	54	27	2
logit_lens_surprisal	log_p_layers	none	188	16	17
mc_word_entropy	log_p_samples_nats, mc_word_entropy	alpha=0.5, 1; n_samples=64, 512; max_tokens=20; seed=0	312	56	2
mmc_unigram_surprisal	log_p_observed, unigram_surprisal	none	55	36	2
mmc_unigram_surprisal_pool	log_p_observed, unigram_surprisal	n_samples=4e+07; seed=900000	13	13	1
mmc_unigram_surprisal_pool_ee	log_p_observed, unigram_surprisal	n_samples=4e+07; seed=900000	10	10	1
modelblocks_surprisal	log_p_observed	none	28	28	1
next_word_entropy	next_word_entropy	none	126	58	2
next_word_renyi_entropy	next_word_renyi_entropy_alpha0.5	alpha=0.5	101	58	2
responsivity	decontextualised_distance, expected_decontextualised_distance	n_samples=512; seed=0	58	29	2
similarity_surprisal_contextual	expected_similarity_tokens, next_word_information_value, next_word_information_value_first, similarity_adjusted_surprisal, similarity_adjusted_surprisal_first	n_samples=50; seed=0	6	6	1
similarity_surprisal_noncontextual	expected_similarity_tokens, next_word_information_value, next_word_information_value_first, similarity_adjusted_surprisal, similarity_adjusted_surprisal_first	none	19	19	1
similarity_surprisal_word	next_word_information_value_orth, next_word_information_value_pos, similarity_adjusted_surprisal_orth, similarity_adjusted_surprisal_pos	n_samples=50; max_tokens=10; seed=0	6	6	1
surprisal	log_p_observed	none	442	57	15
token_renyi_entropy	token_renyi_entropy_alpha0.5, token_renyi_entropy_alpha2.0	alpha=0.5, 2	170	58	2
tuned_lens_surprisal	log_p_layers	none	16	4	4
unigram_surprisal	log_p_observed, unigram_surprisal	none	76	56	5
word_information_value	25 keys; see predictor table or tags.parquet	n_samples=5; max_tokens=3; seed=0	41	41	1
wordsprobability_surprisal	log_p_observed	none	34	25	2
wordsprobability_surprisal_buggy	log_p_observed	none	34	25	2

What each one computes for a word w with left context c, and the piece of the tag it appears as:

method	tag piece	definition
surprisal	surprisal	log P(w | c) from a causal LM, summed over the word's subword tokens
wordsprobability_surprisal	wordsprobability_surprisal	word log-probability with the word-boundary and sentence-start correction of the upstream wordsprobability package
wordsprobability_surprisal_buggy	wordsprobability_surprisal_buggy	the same package's uncorrected column: a plain sum of subword log-probabilities, equal to surprisal aggregated to whitespace words
char_beam	charbeam_surprisal_K<k>	character-level log-probability, marginalising the top-K tokenisations of the string with a byte-level beam search (genlm-bytes), then aggregated to the corpus's units
first_token_entropy	ft_entropy	Shannon entropy of the next-token distribution over the full subword vocabulary, read at the word's first subword
token_renyi_entropy	token_renyi_entropy_a<x>	Renyi-alpha entropy of that same first-subword distribution
next_word_entropy	next_word_entropy	Shannon entropy after restricting the next-token distribution to word-initial tokens and renormalising
next_word_renyi_entropy	next_word_renyi_entropy_a<x>	Renyi-alpha entropy of that restricted distribution
mc_word_entropy	mcword_entropy_N<n>_M<m>_a<x>_s<s>	word entropy by Monte Carlo: N samples, each continued subword by subword until a word boundary is drawn (at most M steps), aggregated at Renyi-alpha
continuation_entropy	cont_entropy_N<n>_L<l>_s<s>	sequence entropy by Monte Carlo: N sampled continuations of up to L tokens each
forward_word_lookahead_entropy	fwd_word_lookahead_entropy_N<n>_M<m>_K<k>_s<s>	look-ahead entropy over whole words at every depth 1..K, in joint, conditional and marginal form
unigram_surprisal	unigram_surprisal	context-free word probability read off a frequency table
mmc_unigram_surprisal	mmc_unigram_surprisal	context-free word probability recovered from an LM by chain-rule Rao-Blackwellisation over sampled contexts
responsivity	responsivity_N<n>_s<s>	mean cosine distance in the model's input-embedding space from the observed word to N sampled alternatives, and among those alternatives

Models and frequency tables

model	model_kind	methods	cells
owt	frequency_table	1	5
wordfreq	frequency_table	1	56
wordfreq-da	frequency_table	1	1
wordfreq-en	frequency_table	1	1
EleutherAI/pythia-1.4b-deduped	lm	2	43
EleutherAI/pythia-12b-deduped	lm	1	4
EleutherAI/pythia-14m	lm	2	39
EleutherAI/pythia-160m-deduped	lm	2	43
EleutherAI/pythia-1b-deduped	lm	1	16
EleutherAI/pythia-2.8b-deduped	lm	1	4
EleutherAI/pythia-31m	lm	2	39
EleutherAI/pythia-410m-deduped	lm	2	43
EleutherAI/pythia-6.9b-deduped	lm	1	4
EleutherAI/pythia-70m-deduped	lm	2	43
Qwen/Qwen3-0.6B	lm	1	29
Qwen/Qwen3-1.7B	lm	1	29
Qwen/Qwen3-4B	lm	1	29
ai-forever/mGPT	lm	17	897
facebook/opt-1.3b	lm	1	4
facebook/opt-125m	lm	3	35
facebook/opt-6.7b	lm	1	4
gpt2	lm	21	623
gpt2-large	lm	3	34
gpt2-medium	lm	2	26
gpt2-xl	lm	3	47
wt103	lm	1	13

wordfreq resolves its language from the corpus; wordfreq-en and wordfreq-da pin it. owt is a whitespace-token count table over OpenWebText, published here as owt/unigram_counts.tsv.

Values from different models are on different scales and different vocabularies. Compare within a model, not across models.

Corpora

EEG (N400) (8): federmeier2007_n400, hubbard2019_n400, michaelov_n400, szewczyk2022_n400, szewczyk_federmeier2022_n400, szewczyk_n400, wlotko_federmeier2012_n400, zuco_n400

eye tracking (45): bsc, celer, copco, devarda2023, dundee, emtec, geco, meco_de, meco_du, meco_ee, meco_en, meco_fi, meco_gr, meco_he, meco_it, meco_ko, meco_l1_w2_ba, meco_l1_w2_bp, meco_l1_w2_ch_s, meco_l1_w2_ch_t, meco_l1_w2_da, meco_l1_w2_en_uk, meco_l1_w2_ge_po, meco_l1_w2_ge_zu, meco_l1_w2_hi_iiith, meco_l1_w2_hi_iitk, meco_l1_w2_ic, meco_l1_w2_no, meco_l1_w2_ru_mo, meco_l1_w2_se, meco_l1_w2_sp_ch, meco_l1_w2_tr, meco_l2_w1, meco_l2_w2, meco_no, meco_ru, meco_sp, meco_tr, onestop, potec, provo, sbsat, ucl_et, zuco1, zuco2

maze (1): natural_stories_maze

self-paced reading (4): brothers_kuperberg, brown_spr, natural_stories, ucl_spr

Files

path	files	cells	rows
(root)	58	1827	121,571,006
cont_entropy_samples/	170	170	3,211,755
fwd_word_lookahead_entropy_samples/	22	22	700,591
lens_layers/	204	204	6,492,037
mcword_samples/	312	312	7,057,470
next_log_probs/	21	21	62,355
responsivity/	58	58	2,463,456
simadj_esim/	25	25	606,073

path	contents
tags.parquet	the cell index described above
history.parquet	one row per file ever published: the commit and time that first added it and that last changed it, and whether it is still present
manifests/<tag>.json	the runner's full provenance sidecar for cells computed since 2026-09-06: resolved options, model snapshot, corpus hash, source-tree hash, package versions, and the environment (commit, device, CUDA build, time)
<dataset>.parquet	scalar per-word predictors for one corpus, one file per corpus
responsivity/	embedding-distance predictors, same long schema, one file per (model, corpus). Has its own README
next_log_probs/	the model's predictive distribution over word types at each scored position
mcword_samples/	the per-word Monte-Carlo sample pool behind mc_word_entropy
cont_entropy_samples/	the per-word sample pool behind continuation_entropy
fwd_word_lookahead_entropy_samples/	word-level look-ahead entropies at depths 1..K, plus the raw per-sample step log-probabilities
owt/unigram_counts.tsv	OpenWebText unigram count table
modelblocks_kenlm_models/	order-1 KenLM ARPA trained on OpenWebText
kenlm_o1/	per-word OpenWebText unigram surprisal, one pickle per corpus

Only tags.parquet and the per-corpus files are dataset-viewer configs. A row holding a vocabulary-sized array does not render usefully, so the array folders are loaded by path.

Loading

One corpus, one cell:

import pandas as pd

REPO = "hf://datasets/replicaverse/replica-predictors"
provo = pd.read_parquet(f"{REPO}/provo.parquet")
cell = provo[provo.tag == "gpt2_surprisal_provo"]
cell[["stimulus_id", "unit_index", "text_unit", "value", "value_unit"]].head()

One predictor per column, ready to model:

wide = (provo[provo.tag.isin(["gpt2_surprisal_provo", "gpt2_ft_entropy_provo"])]
        .pivot_table(index=["stimulus_id", "unit_index", "text_unit"],
                     columns="predictor", values="value")
        .reset_index())

Through the dataset viewer:

from datasets import load_dataset
idx = load_dataset("replicaverse/replica-predictors", "tags", split="train")
provo = load_dataset("replicaverse/replica-predictors", "provo", split="train")

Raw distributions and sample pools, by path:

import numpy as np, pyarrow.parquet as pq

lp = pq.read_table(f"{REPO}/next_log_probs/gpt2_ft_entropy_provo.parquet")
print(lp.schema.field("logprobs").metadata)          # {b'unit': b'nats'}
dist = np.asarray(lp["logprobs"][0].as_py())         # (n_word_types,) log-probs
np.exp(dist).sum()                                   # 1, to float32 rounding

la = pq.read_table(f"{REPO}/fwd_word_lookahead_entropy_samples/"
                   "ai-forever_mGPT_fwd_word_lookahead_entropy_N512_M30_K5_s0_meco_de.parquet")
row = la.slice(0, 1).to_pylist()[0]
steps = np.asarray(row["log_p_step_nats"]).reshape(row["n_samples"], row["k_words"])

Data notes

Text units. unit_index indexes the corpus's own unit layer, which is words everywhere except the twelve meco_char_* corpora, where it is single characters (including spaces). text_unit is that layer's surface string, so for a multi-word interest area it is the whole area, with internal whitespace written as _.

Missing text units. A stimulus whose stored unit count disagrees with the corpus was scored against a different version of that text, so no position in it can be named and every one of its rows gets text_unit = null. Six brothers_kuperberg stimuli are in that state: bk_009_{H,L,M} and bk_058_{H,L,M}, in all of that corpus's cells. Their values do not correspond to the shipped sentences, so filter them out rather than joining them.

Stimulus ids. MECO first-wave rows previously carried a float-formatted id (meco_en_1.0). They are now written as meco_en_1, matching the corpus.

Distributions. next_log_probs/ holds the model's next-position distribution after marginalising the subword vocabulary onto unique surface forms and renormalising, so a row is one log-probability per word type: 32,761 entries for gpt2 and 79,265 for ai-forever/mGPT. Each row sums to 1 under exp, to float32 rounding. It re-derives any distributional statistic over word types offline (entropy, Renyi at any order, top-k mass, the observed word's surprisal).

Precision. Scalar predictors are float64 as computed. The stored distributions and sample pools are float32, so a statistic re-derived from them is not bit-identical to one computed in float64.

Mixed-unit files. A row of fwd_word_lookahead_entropy_samples/ carries six array columns rather than one value, and they are not all in the same unit:

column	unit	contents
joint_word_entropy	bits	H of the next k words jointly, for k = 1..K
conditional_word_entropy	bits	the step from k-1 to k, same length
marginal_word_entropy	bits	H of the word at depth k alone
marginal_word_entropy_plugin	bits	the same without the small-sample correction
log_p_step_nats	nats	raw per-sample step log-probabilities, row-major [n_samples, k_words]
id_hash	identifier	word-id hashes for those samples, same shape

Its per-row value_unit reports bits, the unit of the entropy columns, and tags.parquet names joint_word_entropy as the cell's predictor. The exact unit of every column is in that column's Parquet field metadata, so read that rather than the row when you use one of the other columns.

Coverage. Not every (model, method, corpus) combination exists. Check tags.parquet rather than assuming a cell is present.