README.md

---
pretty_name: "SchLAIMS: A Dataset for Claim Selection in Scholarly Debates"
language:
- en
license: other
task_categories:
- text-classification
size_categories:
- 10K<n<100K
configs:
- config_name: full
  data_files:
  - split: train
    path: data/full/train.parquet
  - split: validation
    path: data/full/dev.parquet
  - split: test
    path: data/full/test.parquet
- config_name: majority
  data_files:
  - split: train
    path: data/majority/train.parquet
  - split: validation
    path: data/majority/dev.parquet
  - split: test
    path: data/majority/test.parquet
---

# SchLAIMS: A Dataset for Claim Selection in Scholarly Debates

## What It Is

SchLAIMS is a dataset for **paragraph-level claim selection**. Given
one scientific paragraph, the task is to select the one sentence that states
the paragraph's central point, or return `null` when the paragraph makes no
single claim.

The release contains 500 linked article, comment, and reply sets found through
OpenAlex: 29,703 paragraphs and 114,138 sentences. In the files, those three
document turns are named `original`, `critique`, and `response`.

## What A Label Means

A claim is the sentence that states the central point the author wants the
reader to accept. Evidence, methods, citations, details, qualifications, and
caveats usually serve as premises supporting that point. Annotators read the
whole paragraph, choose its highest-level takeaway, and do not force a claim
when no sentence clearly plays that role. Each valid label is therefore one
source `sentence_n` or `null`.

One row contains a paragraph, its ordered sentences, its selected claim, its
document turn, and IDs that link it back to its triad and source document.

## Example

This unanimous training example is recoverable in either configuration:

- `triad_name`: `triad_4498`
- `triad_id`: `W3135045181_W4283257629_W4283256895`
- `turn`: `original`
- `paragraph_id`: `triad_4498__original__p0037`
- `paragraph_index`: `37`
- OpenAlex work: `W3135045181`
- Source: Laroche and L'Esperance (2021), *Cancer Incidence and Mortality among Firefighters: An Overview of Epidemiologic Systematic Reviews*, DOI `10.3390/ijerph18052519` (CC BY)

> **1.** The extent to which a systematic review can draw conclusions about the effects of an intervention depends on the validity of the data and results from the included studies.
>
> **2.** In particular, a meta-analysis of invalid or low-quality studies may produce a misleading result, yielding a narrow confidence interval around the wrong intervention effect estimate [16].
>
> **3.** Variations in study quality can explain differences in the findings of studies that are included in a systematic review.
>
> **4.** As a result, the quality of a study will affect the strength of the evidence that can be drawn from it.
>
> **5.** In other words, it determines whether we can be confident that the results of a study reflect the 'truth' and by extrapolation, whether we can be confident in the results of the systematic review [16,17].

The selected claim is sentence 1. All three label passes selected it. The
alternative would be `null` if none of the five sentences expressed one clear
central claim.

## How Labels Were Made

These are AI-consensus reference labels, not human-adjudicated gold labels.
**We used agents to create the training labels only after a hidden test showed
that they could follow the human-written guidelines and usually select the same
claims as human annotators.** The agents never saw the human answers. Across 10
seeded low-reasoning GPT-5.4 mini runs, the retained human-consensus test had
about 0.84 mean claim F1 and 0.948 mean sentence accuracy (161 sentences and 27
claims). This validates the labeling method; it does not mean every released
paragraph was checked by a person. Details are in the
[SchLAIMS validation pipeline](https://github.com/danielscottsmith/SchLAIMS/tree/main/01_agent_validation).

The release applies revised long-block paragraphing, scispaCy sentence
segmentation, and three independent GPT label passes. An exact choice made by
at least two passes becomes the consensus label. The `full` view retains all
paragraphs and deterministically selects one observed vote for each of 337
three-way disagreements. The `majority` view excludes those disagreements. It
is a different, less ambiguous population, not simply a better benchmark.

## Model Results And Human Validation

The released Longformer selector has two jointly reported fixed-test results.
It was also checked on the retained human-consensus paragraphs after removing
one paragraph found verbatim in its training data.

| Reference labels | Population | Paragraphs | Precision | Recall | Claim F1 |
|---|---|---:|---:|---:|---:|
| AI consensus, full test | Fixed test | 4,603 | 0.7032 | 0.7039 | **0.7035** |
| AI consensus, majority-only test | Fixed test | 4,538 | 0.7117 | 0.7106 | **0.7111** |
| Human consensus | External validation | 26 | 0.7692 | 0.7692 | **0.7692** |

The human row is a small external validation check, not a third canonical test
benchmark. Its source table, exclusions, overlap audit, and hashes are provided
as auxiliary validation provenance. The complete 34-paragraph diagnostic is
not a reported model result.

## Splits And Loading

The split contains 350 training, 75 development, and 75 test triads. Linked
article, comment, and reply documents always stay together.

```python
from datasets import load_dataset

full = load_dataset("danielscottsmith/schlaims", "full")
majority = load_dataset("danielscottsmith/schlaims", "majority")
```

Native Parquet and JSONL files support ordinary use. The three hashed archives preserve
raw sources, exact processed data, and label provenance for reproduction.

## Uses And Limits

The dataset supports training and evaluating paragraph-level claim selectors,
studying argumentative structure, and assisting scientific-text annotation.
It is not fact checking, evidence-quality grading, importance ranking,
extraction of every possible claim, or isolated-sentence classification. These
500 triads do not represent science as a whole.

## Rights And Citation

Source prose remains under its original publisher terms and is redistributed
under the PI-approved basis for this release. OpenAlex's CC0 metadata license
does not relicense article text. Project-created labels, boundaries, audits,
and manifests are available under CC BY 4.0.

```bibtex
@dataset{smith_schlaims_2026,
  author  = {Smith, D. S. and Verdi, D. A. and Chen, R. and Zhang, H. and McFarland, D. A.},
  title   = {SchLAIMS: A Dataset for Claim Selection in Scholarly Debates},
  year    = {2026},
  version = {1},
  url     = {https://huggingface.co/datasets/danielscottsmith/schlaims}
}
```

Please also cite OpenAlex: Priem et al. (2022), "OpenAlex: A fully-open index
of scholarly works, authors, venues, institutions, and concepts," *arXiv:2205.01833*.