union-len64-130M

Procedurally generated set-union task examples (Procedural Pretraining, Jiang et al. 2026), as flat uint16 token-id .bin files. Each row presents two random symbol sequences and asks for their union — the distinct symbols in order of first appearance — laid out as s1 | SEP | s2 | SEP | union(padded). Symbols are drawn from [1, 100); SEP = 100 and PAD = 101, so ids span [0, 102) and the vocabulary is 102.

Task parameters

param	value
n_symbols	100
input_len (combined, split into two halves)	64
seq_length (row = 2·input_len + 2)	130

file	split	tokens
train.bin	train	130,000,000
val.bin	val	2,600,000

The bin is a concatenation of fixed-length 130-token rows (num_tokens is a whole multiple of it): load via tokens.reshape(-1, 130) to train one example per row. The layout is fixed, so the union output occupies the tail from position 66 onward — downstream training masks the loss to that tail. train (seed 0) and val (a disjoint generator seed) are independent streams of the same task; token count = filesize / 2; train.meta.json / val.meta.json carry the full config.

Load a bin with the standard Hugging Face downloader:

from huggingface_hub import hf_hub_download
import numpy as np

path = hf_hub_download(repo_id="alexkstern/union-len64-130M", filename="train.bin", repo_type="dataset")
tokens = np.memmap(path, dtype="uint16", mode="r")