---
Language
hi
en
License
cc-by-4.0
Task categories
text-classification
Tags
natural-language-inference
fact-verification
citation-verification
hinglish
code-mixing
rag
Pretty name
HingVerify
Size categories
1K<n<10K
Configs
- config_name: nli
  data_files:
  - split: train
    path: train.jsonl
  - split: validation
    path: validation.jsonl
  - split: test
    path: test.jsonl
- config_name: corpus
  data_files:
  - split: train
    path: corpus.jsonl
- config_name: rag_questions
  data_files:
  - split: test
    path: rag_questions.jsonl
HingVerify: citation verification in English, Hindi and Hinglish
Claim–evidence pairs for checking whether a sentence is supported by the official passage it cites, built from official pages (PIB, PMIndia, pmkisan.gov.in) about Indian central government welfare schemes. Claims are in English, Hindi (Devanagari) and Hinglish (Roman-script Hindi mixed with English); evidence is in English or Hindi.

File	Rows	What it is
test.jsonl	366	Human-adjudicated test set from 6 schemes (PM-KISAN, PM-JAY, PMUY, PMAY-G, SSY, APY). Frozen on 2026-10-21, SHA-256 ab83912a81e40f9d8d2d952ea566398da01a69fabb51e5d083ac8f2d3f00d3b9
train.jsonl	801	Synthetic training pairs from 14 other schemes (LLM-written hard negatives)
validation.jsonl	126	Synthetic validation pairs from 3 further schemes
corpus.jsonl	159	All official passages (train, validation and test schemes), each with its source URL
rag_questions.jsonl	60	Citizen questions (20 per language, 12 unanswerable) with gold passage and reference answer
Schemes are split so that no test scheme appears in training or validation.

Labels
SUPPORTED, CONTRADICTED, NOT_ENOUGH_INFO (the passage neither confirms nor rules out the claim). label_binary collapses the last two into NOT_SUPPORTED. error_type names how the claim was built (e.g. wrong_number, negation, unit_scale, paraphrase). source is drafted or team (written by the authors) in the test set and synthetic in train/validation (all written by Claude, 24 Sep 2026); crosslingual is true when claim and evidence are in different languages (a Hinglish claim with Hindi evidence counts as same-language).

How the test set was made
306 claims were drafted by an LLM (Claude, Anthropic), 34 per language–label cell, and 60 were written by the authors. Two authors (annotator A) labelled the 306 drafted claims, 153 each, without seeing the draft labels or any other label; for the 60 hand-written claims the author's intended label served as A. A separate LLM session with no access to the draft labels (annotator B) labelled all 366 pairs, and a third author adjudicated the 20 pairs where A and B disagreed. Cohen's kappa between the human annotator and the final labels on the drafted claims is 0.946 (A vs B: 0.918). Because the drafts and the LLM annotator share a model family, and the final labels derive from annotator A, two independent people were also compared: after the freeze each annotator relabelled 25 pairs from the other's half (stratified, no labels shown), giving 82% agreement and kappa 0.729 (95% bootstrap CI 0.55-0.88). Most disagreements were CONTRADICTED versus NOT_ENOUGH_INFO on missing-information claims. The frozen labels were not changed.

Use
from datasets import load_dataset
nli = load_dataset("Teena2812/hingverify", "nli")
print(nli["test"][0])
Please do not train on test.jsonl; it is the evaluation set reported in the paper.

Limitations
Training and validation data are synthetic, the domain is one country's welfare schemes, and scheme figures change over time (each passage gives its source URL). A label says whether the passage supports the claim, not whether the claim is true today.

Model and paper
Fine-tuned verifier: Teena2812/hingverify-mdeberta. Paper: A. Singh, T. Kalia, M. Malik and S. Gupta, "HingVerify: A Compact Fine-Tuned NLI Verifier for Citation Grounding in Hindi and Hinglish Retrieval-Augmented Generation" (under review).

License
CC BY 4.0 for the claims, labels and questions. Evidence passages are quoted from Government of India web pages, which remain subject to their own terms; each row links to its source.