ChaosNLI v1.0 — community preservation mirror

This repository is an unofficial preservation mirror of ChaosNLI v1.0, introduced by Yixin Nie, Xiang Zhou and Mohit Bansal at EMNLP 2020. The archive was obtained from a Kaggle mirror. The archive and extracted source files are preserved without modification. This repository adds documentation, checksums, validation code and Hugging Face loading configuration; it does not add annotations or revise labels. Please cite the original paper.

The mirror is not affiliated with or endorsed by the original authors. The Kaggle uploader (curiosityquotient/chaos-nli) is credited as the intermediary from which the archive was obtained, not as an author. The maintainer's contribution is preservation and packaging.

Preservation release mirror-v1.0.0. Verified data revision: eb7cc18dbfa0f161d93a2b22c7856ca0faafcf2b. The preserved payloads at that revision were downloaded anonymously and checked on 2026-09-26; see Verification. Pin the full data revision in research code; the tag is a convenience name.

Contents

ChaosNLI provides 100 new crowd annotations per example for examples drawn from the development sets of SNLI, MultiNLI (matched) and αNLI, with the counts and distributions stored in each record. It is not a file of 100 individual annotation records per example, and it identifies no annotators. [Nie et al., 2020, section 3]

Configuration	Raw file	Task	Records	New votes	Vector order
snli	raw/chaosNLI_snli.jsonl	three-way NLI	1,514	151,400	[e, n, c]
mnli_m	raw/chaosNLI_mnli_m.jsonl	three-way NLI	1,599	159,900	[e, n, c]
alphanli	raw/chaosNLI_alphanli.jsonl	two-choice abductive NLI	1,532	153,200	[1, 2]
total			4,645	464,500

The two three-way components together contain 3,113 records. Counts are taken from the full-record audit.

Each configuration has a single split named validation. This is the mirror's loader name for the existing development-set material; it is not a newly sampled validation split. There is no train or test split, and there is no combined configuration: αNLI is a two-choice task and is kept separate from three-way NLI.

Repository layout

Path	What it is
archive/kaggle-download.zip	The ZIP exactly as downloaded from Kaggle (562,342 bytes)
raw/	The four archive members, byte-for-byte: README.txt and the three JSONL files
LICENSE	The ChaosNLI licence file, copied unchanged from the authors' repository
upstream/ChaosNLI/	Commit-pinned copies of the authors' README.md, LICENSE and download_data.sh (never executed)
SHA256SUMS	Digests of the archive, the raw files, the root licence and the upstream snapshots only
audit/	Validation results, anomaly identifiers, source manifest, claim map and execution receipts
scripts/, tests/, schemas/, sources/	Validation, packaging and verification code
PROVENANCE.md, THIRD_PARTY_NOTICES.md, REPRODUCIBILITY.md, MAINTENANCE.md, CHANGELOG.md, CITATION.bib	Supporting documentation

The archive's own raw/README.txt is the authors' document and is preserved unchanged, including its old repository URL. This README.md is the mirror's dataset card; neither replaces the other.

Usage

Each example names a configuration and pins the verified data revision. No custom loading code and no trust_remote_code are needed.

from datasets import load_dataset

DATA_REVISION = "eb7cc18dbfa0f161d93a2b22c7856ca0faafcf2b"

snli = load_dataset(
    "earino/chaosnli", "snli",
    split="validation", revision=DATA_REVISION,
)
mnli = load_dataset(
    "earino/chaosnli", "mnli_m",
    split="validation", revision=DATA_REVISION,
)
alpha = load_dataset(
    "earino/chaosnli", "alphanli",
    split="validation", revision=DATA_REVISION,
)

Download the untouched archive:

from huggingface_hub import hf_hub_download

zip_path = hf_hub_download(
    repo_id="earino/chaosnli",
    filename="archive/kaggle-download.zip",
    repo_type="dataset",
    revision="eb7cc18dbfa0f161d93a2b22c7856ca0faafcf2b",
)

Direct links to the preserved files at the verified revision: archive/kaggle-download.zip, raw/README.txt, raw/chaosNLI_snli.jsonl, raw/chaosNLI_mnli_m.jsonl, raw/chaosNLI_alphanli.jsonl.

Verifying the files

SHA256SUMS covers the preserved archive, the four raw files, the root LICENSE and the three upstream snapshots. It does not cover the documentation, code or audit files; identify those by the repository commit. From a downloaded copy of the repository:

sha256sum -c SHA256SUMS        # Linux
shasum -a 256 -c SHA256SUMS    # macOS

The archive digest is 419c0dcf4ae1f4153654b7927880b6c7fde7485568f9efa4540626748c2e889a. To rerun the full validation and compare it with the committed audit, see REPRODUCIBILITY.md.

Data fields

All three files are JSON Lines, one object per line, in the authors' format. [Authors' README]

Field	Type	Meaning
uid	string	Example identifier. For SNLI and MNLI it is the original pairID; for αNLI the original story_id.
label_counter	object	New ChaosNLI vote counts keyed by label (e, n, c or "1", "2"). A label with zero votes may be absent from this object in the raw files.
majority_label	string (NLI) or integer (αNLI)	The majority of the new votes, as supplied by the authors.
label_dist	list of floats	New vote proportions in the fixed vector order.
label_count	list of integers	New vote counts in the fixed vector order; each sums to 100.
entropy	float	Entropy of label_dist in bits, as supplied.
example.uid	string	Same as uid.
example.premise, example.hypothesis	string	SNLI and MNLI text pair.
example.obs1, example.obs2, example.hyp1, example.hyp2	string	αNLI beginning and ending observations and the two candidate hypotheses.
example.source	string	Source tag: snli_agree_3, mnli_agree_3 or abdnli_dev.
old_label	string (NLI) or integer (αNLI)	The label in the original dataset (e, n, c, or 1/2).
old_labels	list of strings	SNLI and MNLI only: the original individual annotations (entailment, neutral, contradiction). αNLI has no such field.

Vector order. For SNLI and MNLI, label_count and label_dist are ordered [entailment, neutral, contradiction]. For αNLI they are ordered [hypothesis 1, hypothesis 2]. Do not derive the order from label_counter key order, which varies between records.

Old and new labels. old_label and old_labels come from the original datasets; label_count, label_dist, label_counter, majority_label and entropy describe the new ChaosNLI votes. They need not agree: the supplied old_label differs from the new majority_label in 378 SNLI, 508 MNLI and 163 αNLI records.

Ties. 14 SNLI, 14 MNLI and 8 αNLI records have two labels tied for the highest new vote count. The supplied majority_label is one of the tied labels and is kept as released; the mirror does not break ties differently. Identifiers are listed in the audit.

Short old-annotation lists. 7 SNLI records have four entries in old_labels instead of five. They still have 100 new votes each. The same four annotations appear in the original SNLI 1.0 development file, so nothing is missing from the ChaosNLI release; no annotation has been inserted.

Sparse label_counter. 546 SNLI, 174 MNLI and 114 αNLI records omit a zero-count label from label_counter. label_count still has an explicit zero. The raw files keep the omission.

Loaded representation

The configurations use the standard JSON loader with explicit typed features declared in this card. With the pinned datasets 5.0.1, loading behaves as follows:

• label_counter is a struct with one field per label, so a key absent from the raw object loads as null. A null here means zero votes, not a missing annotation.
• All other values, including every entropy and label_dist float, load unchanged. The local and remote round-trip tests compare every loaded row with an independent parse of the raw file.
• Without the typed features, the same datasets version loads label_counter as a JSON-encoded field and re-serialises each line with ten-digit float precision. In the recorded probe this changed 1,493 of 1,514 SNLI entropy values (loader audit). The typed features avoid that path. The raw JSONL is never rewritten to suit a loader.

Byte-level fidelity is established by comparing the downloaded files, not by the parsed representation.

Source and purpose

ChaosNLI was built to study the distribution of human opinions on NLI examples, rather than a single majority label, and to test how well models recover that distribution. [Nie et al., 2020]

According to the paper (section 3.1), the SNLI and MNLI components are subsets of the SNLI development set and the MNLI-matched development set whose original majority label agreed with only three of the five original annotations; the αNLI component is the entire αNLI development set. The new annotations were collected from qualified Amazon Mechanical Turk workers (section 3.2).

The authors separately distributed saved model predictions. Those predictions are not included here; they are not needed to preserve these dataset files, and this mirror does not reproduce the authors' model results.

Provenance

The owner downloaded archive.zip from Kaggle. The download date was not recorded. The archive and its members are published unchanged; the member paths chaosNLI_v1.0/* are mapped to raw/*. The authors' documented Dropbox link returned an HTML page titled "File Deleted" when tested on 2026-09-26.

The checks establish:

1. Local integrity: the archive, member CRCs and member digests; every raw file equals its archive member byte-for-byte.
2. Consistency with the release documentation: file names, record counts, the vote total, fields and vector order match the authors' README, the archive's README.txt and the paper.
3. Independent content comparison: text fields and original labels of all 4,645 records match the original SNLI 1.0, MultiNLI 1.0 and AI2 αNLI development files.
4. Hosted copy: anonymous download and loading of this repository (see Verification).

No author-issued archive or author-published checksum was available, so the mirror does not claim that its archive is byte-identical to the authors' original distribution package. Full details: PROVENANCE.md and audit/source-manifest.json.

Licence

The ChaosNLI authors distribute ChaosNLI under the Creative Commons Attribution-NonCommercial 4.0 International licence (legal code). The exact upstream licence file is kept as LICENSE. The metadata tag cc-by-nc-4.0 reports that statement; it is not a legal assurance that every downstream use is permitted.

The examples come from SNLI, MultiNLI and αNLI, which carry their own notices; for instance, the SNLI page states a Creative Commons Attribution-ShareAlike 4.0 licence. The reannotation does not remove rights in the underlying texts. See THIRD_PARTY_NOTICES.md.

The documentation, code and audit files added by this mirror are offered under the same CC BY-NC 4.0 terms. No access gate, payment or additional terms apply.

Limitations

• The new votes describe the crowd workers who took part in the original study, under its instructions and quality controls. They are not a sample of any other population.
• A majority label, old or new, is not proof of an objective truth; many of these examples were chosen because annotators disagreed.
• The SNLI and MNLI components were deliberately selected low-agreement development examples. None of the components is a random sample of production inputs.
• These examples have been public since 2015–2020. This mirror does not measure whether later models were exposed to them during training.
• The audit checks internal consistency and agreement with documentation and original development sets. It does not certify that the annotations are free of error or bias.

Verification

On 2026-09-26, a fresh process with a new, empty Hugging Face cache, no stored token and credential variables removed downloaded data revision eb7cc18dbfa0f161d93a2b22c7856ca0faafcf2b anonymously and confirmed that:

• the archive digest and all 9 SHA256SUMS entries matched, and each raw file was byte-identical to its archive member;
• the validator and the reproduction check regenerated the committed audit (PASS);
• all three configurations loaded with only the validation split, and all 4,645 rows passed the semantic round-trip comparison (PASS);
• the Dataset Viewer reported the three configurations with 1,514, 1,599 and 1,532 rows.

The report is audit/remote-verification.json; release.json records the data revision and the source archive digest. These checks concern the data revision. The documentation revision that contains this section cannot record a check of itself; pin the data revision above.

Citation

Please cite the original paper:

@inproceedings{nie-etal-2020-learn,
    title = "What Can We Learn from Collective Human Opinions on Natural Language Inference Data?",
    author = "Nie, Yixin  and
      Zhou, Xiang  and
      Bansal, Mohit",
    editor = "Webber, Bonnie  and
      Cohn, Trevor  and
      He, Yulan  and
      Liu, Yang",
    booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
    month = nov,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2020.emnlp-main.734/",
    doi = "10.18653/v1/2020.emnlp-main.734",
    pages = "9131--9143"
}

When using the underlying examples, also credit SNLI (Bowman et al., 2015), MultiNLI (Williams et al., 2018) and αNLI (Bhagavatula et al., 2020); see CITATION.bib.

Assistance disclosure

This mirror was prepared by Claude Code (Anthropic's command-line agent, model Claude Opus 5.5), working in the owner's local environment at the owner's request and from the owner's written specification. The agent wrote the scripts, tests and documentation, ran the checks, and made the Hugging Face uploads with the owner's authenticated account. The evidence consists of automated checks (the validator, the independent recount, the unit tests, and the local and remote loading tests) plus the agent's own review of their output. No independent human audit of the data or code has been performed.

Maintenance

See MAINTENANCE.md and CHANGELOG.md. The raw v1.0 bytes will not be replaced under the same paths. This is one additional, verifiable copy; no hosted copy is permanent.