AXN:0BAE.CAPTURE

AXN:0BAE.CAPTURE.🔽🔩⏪🪐🔃💚

Observation

Google AI Overview, 2026-11-01 · the capture

Description

THE CORPUS, SIGNED OUT, 2026-11-01: 'Dolma is an open-source 3-trillion-token English corpus created by the Allen Institute for Artificial Intelligence (Ai2) to support language model pretraining and transparency.' Access (Hugging Face, the GitHub toolkit); key features: scale ('version 1.7 utilizes about 1.715 trillion actual training tokens across ~5.2 billion documents totaling over 5,300 GB'), data sources by kind with their named sets, the curation pipeline ('Bloom filter deduplication (by URL, paragraph, and document), PII/toxicity masking'), licensing ('ODC-BY and Ai2's ImpACT license'), primary use (OLMo); two follow-up questions. Six cards: Hugging Face, the GitHub Pages documentation, Ai2, arXiv, Deep (Learning) Focus, ACL Anthology.

Reading

The innocent reading first. Asked for the dataset, the Overview says what it is, who made it and why, where to get it, what it holds, how it was cleaned, under what licences it is released and which models it trains, from the dataset's own pages and its paper.

Analysis

The corpus composed accurately and more finely than the saved abstract, from the dataset's own pages, with Ai2 as maker where the census row names only the paper's authors. Whether a document in the corpus keeps its author's name, the census row's open question, is not reached.

Match

BROAD MATCH — 'DOLMA DATASET' ON THE OVERVIEW — THE CORPUS AT ITS GRAIN AND PAST THE SAVED ABSTRACT; AI2 NAMED WHERE THE CENSUS NAMES NO INSTITUTION; WHETHER A DOCUMENT KEEPS ITS AUTHOR NOT CARRIED

PER

0.25 — Ai2 named and no author; arXiv 2402.00159 on a card; the dataset card, documentation, Ai2 page and paper carded.

Citations

6 (0 archive-controlled)

Sealed observation

/data/captures/sealed/dolma-dataset-aio-20261101.json — SHA-256 d531d70ddecc85a7f6057ab6d48ff021a7c1a607276bb6c4371d66dc843e68a1