jsanzolac/glove_2024_wikigiga_300d

glove.2024.wikigiga.300d (Stanford, mirrored)

Stanford's 2024 release of GloVe trained on Wikipedia + Gigaword. Source: https://nlp.stanford.edu/data/wordvecs/glove.2024.wikigiga.300d.zip • Training README: https://github.com/stanfordnlp/GloVe/blob/master/Training_README.md.

Vocab size: 1,291,147 (unpruned; whatever Stanford shipped at VOCAB_MIN_COUNT=20)
Dimension: 300
Training tokens: 11.9B
Cased: no (lowercase per -lowerCase flag)
Tokenizer used at training: stanza-en PTBTokenizer -preserveLines -lowerCase -options untokenizable=allKeep

Files

• glove_vectors.bin — raw np.float32, C-order, shape (1291147, 300). Row N is the vector for the word on line N of glove_vocab.txt.
• glove_vocab.txt — one word per line, UTF-8.
• config.json — metadata used by downstream notebooks.

Quick load

import numpy as np
vecs = np.memmap("glove_vectors.bin", dtype=np.float32, mode="r", shape=(1291147, 300))
with open("glove_vocab.txt", encoding="utf-8") as f:
    vocab = [line.rstrip() for line in f]
word2idx = {w: i for i, w in enumerate(vocab)}

Notes

• This is a mirror of Stanford's release, repackaged into a binary + vocab.txt layout that matches jsanzolac/drifting-glove-distilled-r300 so the same training notebooks load it without modification.
• License inherited from the original release (Public Domain Dedication and License, ODC PDDL).