README.md

---
license: cc0-1.0
task_categories:
- text-classification
- text-generation
language:
- en
size_categories:
- 1K<n<10K
tags:
- provenance
- low-background
- data-curation
- dataset-contamination
- dendro
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# Dendro low-background corpus

5034 arXiv records annotated with **archival evidence of when they existed**, produced by
[Dendro](https://github.com/NagaYu/dendro) v0.1.0.

2500 of them (49.7%) are **low-background**: an independent registration record
places them before 2021-01-01, i.e. before large-scale text generation. The name is from
metallurgy — low-background steel is steel smelted before the 1945 atmospheric tests: not
special steel, just ordinary steel that happens to predate the contamination, and valuable
because no amount of later care can reproduce it.

## Read this before you use it

- **This is not an AI-writing detector, and these columns must not be used as one.** They
  report *evidence that content existed before a date*. Nothing here supports a conclusion
  that a particular person did or did not write something.
- **Absence of evidence is not evidence.** Most text that has ever existed was never archived.
  `dendro_abstained` marks rows where the evidence is too thin to act on — those are
  *unknown*, not *generated*.
- **Every row is witnessed by arXiv and arXiv alone.** `dendro_operators` is 1 throughout, and
  `dendro_flags` contains `single_operator`. Compromising one organisation would be sufficient
  to move every date in this dataset. Dendro's whole argument is that confidence should come
  from *independent operators*; this corpus has one, and the column says so rather than
  hiding it.
- **The recent split is "human-attributed", not "verified human".** By 2025 an unknown share
  of abstracts have been through a model. There is no clean recent human corpus to be had —
  which is the argument for dating by archive rather than by inspection.
- **Selection is not neutral.** Archived text over-represents what crawlers reached: English,
  institutional, indexed, long-lived. A low-background subset concentrates every one of those
  biases.

## Columns

| column | meaning |
|---|---|
| `text` | title + abstract, as fingerprinted |
| `title`, `doi` | the record's own metadata |
| `url`, `arxiv_id`, `category`, `era` | identifiers and provenance of the record |
| `published` | arXiv v1 submission date — the witness |
| `dendro_not_after` | date the content is **proven to have existed by** |
| `dendro_not_after_year` | integer year, convenient for filtering |
| `dendro_human_origin_p` | calibrated probability — **not** a label |
| `dendro_ci_low` / `dendro_ci_high` | 90% credible interval, the honest width of the claim |
| `dendro_abstained` | true when the evidence is too thin to act on |
| `dendro_operators` | count of *independent* operators behind the bound |
| `dendro_evidence_logodds` | how hard the bound would be to forge |
| `dendro_flags` | detected inconsistencies |
| `dendro_explanation` | human-readable receipt naming the witness sources |
| `low_background` | proven pre-2021 existence **and** `dendro_ci_low` ≥ 0.8 |

### Why the threshold is 0.8 and not 0.95

Because that is what the evidence in *this* corpus can actually support, and moving the
threshold to flatter the number would be the whole problem with provenance tooling.

Every row here rests on **one operator**. Under Dendro's model a single arXiv registration
gives an evidence log-odds of ~6.5, which puts the *lower* end of the 90% credible interval at
about 0.85 no matter how trustworthy arXiv is — the interval is wide because corroboration is
absent, not because the record is doubtful. At `ci_low ≥ 0.90` this dataset would contain
**zero** low-background rows, which is a true statement about single-operator evidence and a
useless artefact.

So the flag is set where single-operator registration evidence genuinely reaches, and the
number that would tighten it is `dendro_operators` — not the threshold. A document witnessed
by arXiv *and* the Internet Archive *and* Common Crawl clears 0.95 comfortably; filter on
`dendro_ci_low` yourself if you want a different line.

There is deliberately **no** `is_synthetic` column and no boolean verdict. Code that wants a
filter must pick a threshold explicitly, and should prefer `dendro_ci_low` so the choice is
conservative.

## Three lines to use it

```python
from datasets import load_dataset
ds = load_dataset("NagaYu/dendro-lowbackground", split="train")
clean = ds.filter(lambda r: r["low_background"])          # 2500 rows
```

## Purity / retention trade-off

Selecting to bound the *expected* synthetic fraction (provably maximum-cardinality — see
`dendro.corpus_report.build_low_background_subset`):

| max expected synthetic | rows kept | retention |
|---|---|---|
| ≤ 1% | 0 | 0.0% |
| ≤ 2% | 0 | 0.0% |
| ≤ 5% | 0 | 0.0% |
| ≤ 10% | 0 | 0.0% |
| ≤ 20% | 2790 | 55.4% |

## Method, in one paragraph

Dendro asks independent archives — the Internet Archive, Common Crawl, arXiv, Crossref, public
posting archives — what they observed and when, groups witnesses by **operator** rather than by
count (twenty captures from one archive are one archive), and reports the earliest time whose
surviving evidence clears a failure-probability budget. It never reads the prose, which is why
its accuracy does not move when a new generator ships. Probabilities are produced by a mixture
that collapses to the base rate when evidence is absent, so "no evidence" cannot be reported as
"synthetic".

The calibrator is **left unfitted here on purpose**: a fitted map encodes the base rate of the
corpus it was fitted on, and baking that into a published artefact would make every row depend
on a split nobody downstream can inspect. These are the raw numbers.

## Provenance of the data itself

arXiv metadata, including abstracts, is offered under
[CC0 1.0](https://info.arxiv.org/help/api/tou.html) through the public API and OAI-PMH, which
is how it was harvested. Thank you to arXiv for use of its open access interoperability.

Generated 2026-08-10 by `scripts/publish_dataset.py` · Apache-2.0 code, CC0 data.