ichangzii/pit2022-10b

PIT ≤2022 — Point-in-Time Pretraining Corpus (10B tokens)

A lookahead-free pretraining corpus: every source is datable to ≤ 2022-12-31, so a model trained on it has a clean, verifiable knowledge cutoff (no post-2022 contamination). GPT-2 tokenizer, uint16 shards. SmolLM2-style domain mix.

Composition (10B tokens)

Source	Domain	Tokens	Share	≤2022 basis
FineWeb-Edu	Quality web	3.8B	38%	CC dump ≤2022
DCLM	Diverse web	2.5B	25%	~99% ≤2022 (old crawls); drop 2023+ mentions
The Stack (dedup)	Code	1.2B	12%	collected ~2022
Gutenberg	Books	1.0B	10%	public domain (pre-1928)
OpenWebMath	Math	0.8B	8%	row date ≤2022
Wikipedia	Facts	0.7B	7%	2022-03 snapshot

Web 63% · non-web 37%.

Format

• 100 shards named {source}_{NNN}.bin — raw uint16 GPT-2 token IDs (vocab < 65536), EOS-separated docs.
• manifest.json lists all shards + per-source token counts.

Usage

from huggingface_hub import snapshot_download
import numpy as np, json
d = snapshot_download("ichangzii/pit2022-10b", repo_type="dataset")
m = json.load(open(f"{d}/manifest.json"))
shard = np.memmap(f"{d}/{m['shards'][0]}", dtype=np.uint16, mode="r")

Cutoff / provenance

The cutoff is enforced by selecting datable sources (CC dump date, snapshot date, publication date) — not by content heuristics. DCLM lacks a date field but is ~99% ≤2022 by construction (sampled: 100% of URL-dated docs ≤2022, only 0.8% mention 2023+), so it gets only a light regex dropping explicit 2023+ mentions. Built for lookahead-free LM evaluation (e.g. trading backtests).