OLMoTrace

Provenance meta-aggregator / Models and training data

Models and training datatracing toolactive

OLMoTrace is a tracing tool from Allen Institute for AI (Ai2), with the University of Washington, UC Berkeley and Stanford. The object it attaches provenance to: a span of a language model's output and the training documents that contain it verbatim.

Visit OLMoTrace ↗Run an erasure audit

We present OLMoTrace, the first system that traces the outputs of language models back to their full, multi-trillion-token training data in real time.

In its own words: https://arxiv.org/abs/2504.07096, read 2026-10-29

What it carries

What receives provenance	a span of a language model's output and the training documents that contain it verbatim
Relations carried	verbatim match between output and training text op
Carrier	the model's full training corpus, indexed with an extended infini-gram
Expresses · records · tests	expressesrecordstests
Operations tested	model generation, traced back to training text
What survived	returns matching documents within a few seconds; the authors state the matches are not causal and not citations

op the operator's own statement · 3p a third-party account · obs our observation · rd our reading. Hover a tag for its quote; each opens below with its source.

Entity

Canonical name	OLMoTrace
Address	https://playground.allenai.org
Operator	Allen Institute for AI (Ai2), with the University of Washington, UC Berkeley and Stanford
Kind	tracing tool
Unit	a maximal matching span of model output
Lifecycle	active — ACL 2025 demo track
Count	3.2 billion documents and 4.6 trillion tokens (OLMo-2-32B-Instruct's training data) (as of the paper, v2) op
Data access	Ai2 Playground; open source op
Versions and dates	v1: 2025-04-09; v2: 2025-07-08; read: 2026-10-29
Observation	2026-10-29: read

Relations

Neighbours

Data Provenance Initiative

Evidenceevery field's quote, its source and the hash of the saved copy5

self description op
We present OLMoTrace, the first system that traces the outputs of language models back to their full, multi-trillion-token training data in real time.
https://arxiv.org/abs/2504.07096 · read 2026-10-29 · saved copy sha256 cd982680a1ca4c02…
relation types op
OLMoTrace finds and shows verbatim matches between segments of language model output and documents in the training text corpora.
https://arxiv.org/abs/2504.07096 · read 2026-10-29 · saved copy sha256 cd982680a1ca4c02…
limit op
The retrieved documents should not be interpreted as having a causal effect on the LM output, or as supporting evidence or citations for the LM output.
https://arxiv.org/html/2504.07096v2 · read 2026-10-29 · saved copy sha256 b1b16ba583eeb0db…
count op
which totals 3.2 billion documents and 4.6 trillion (Llama-2) tokens
https://arxiv.org/html/2504.07096v2 · read 2026-10-29 · saved copy sha256 b1b16ba583eeb0db…
data access op
OLMoTrace is publicly available and fully open-source.
https://arxiv.org/abs/2504.07096 · read 2026-10-29 · saved copy sha256 cd982680a1ca4c02…

Negative tagswhat it is not3

the retrieved documents are not causes of the output
NOT a citation system
traces only models whose training data is open

Sourcesevery source read, with its date and hash2

https://arxiv.org/abs/2504.07096
read 2026-10-29 · HTTP 200 · sha256 cd982680a1ca4c02…
https://arxiv.org/html/2504.07096v2
read 2026-10-29 · HTTP 200 · sha256 b1b16ba583eeb0db…

Not yet checkedwhat remains to read2

the Playground itself
infini-gram as a separate record

Cite this record

Lee Sharks, ed. "OLMoTrace." Provenance meta-aggregator, Provenance Erasure, compiled 2026-10-29. https://www.provenanceerasure.org/aggregators/olmotrace

Record https://www.provenanceerasure.org/aggregators/olmotrace · entity https://www.provenanceerasure.org/aggregators/olmotrace#entity · the census as data

@misc{pe-olmotrace2026,
  editor = {Sharks, Lee},
  title = {OLMoTrace},
  howpublished = {Provenance meta-aggregator, Provenance Erasure},
  year = {2026},
  note = {Compiled 2026-10-29; every field sourced and dated},
  url = {https://www.provenanceerasure.org/aggregators/olmotrace}
}

Editor: Lee Sharks (ORCID 0009-0000-1599-0703). This page is the provenance meta-aggregator's record of the resource above: its reading and fields are ours, and each field carries its source, its basis and the hash of the copy read. Compiled 2026-10-29. CC BY 4.0.