samuki-hf/psycholing-predictors

Per-word predictors for reading-time modeling

A collection of precomputed per-word predictors for computational models of human sentence processing: surprisal, entropy (Shannon and Renyi), unigram surprisal, and related quantities. One Parquet file per reading corpus.

Schema

column	type	meaning
dataset	string	corpus name (also the file name)
tag	string	cell id (model + method + config), e.g. gpt2_standard_provo
predictor	string	predictor key, e.g. log_p_observed, token_entropy, mc_word_entropy
stimulus_id	string	stimulus id within the dataset
unit_index	int32	0-based word index within the stimulus
value	double	predictor value (NaN where undefined)

Loading

import pandas as pd
provo = pd.read_parquet("hf://datasets/samuki-hf/psycholing-predictors/provo.parquet")
cell = provo[provo.tag == "gpt2_standard_provo"]

Directory layout

path	contents
<dataset>.parquet	scalar per-word predictors, one file per corpus (schema above)
next_log_probs/	full per-word next-token log-prob distributions (float32)
mcword_samples/	Monte-Carlo sample pools behind mc_word_entropy
cont_entropy_samples/	Monte-Carlo sample pools behind continuation_entropy
fwd_word_lookahead_samples/	word-level look-ahead entropy samples (joint / conditional / marginal at depths k=1..K)
owt/unigram_counts.tsv	OpenWebText unigram count table (count-MLE unigram-surprisal baseline)
modelblocks_kenlm_models/	order-1 KenLM ARPA trained on OpenWebText
kenlm_o1/	per-word OpenWebText unigram surprisal, one pickle per dataset

Files under the sample and distribution directories hold list<float> arrays and are not dataset-viewer configs; load them by path with pandas.read_parquet.