Hostility Detection Dataset in Hindi
Paper • 2011.03588 • Published
Datasets:
Dataset Preview
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
The dataset generation failed because of a cast error
Error code: DatasetGenerationCastError
Exception: DatasetGenerationCastError
Message: An error occurred while generating the dataset
All the data files must have the same columns, but at some point there are 7 new columns ({'Reason', 'Category', 'Labels Set', 'split', 'adjudicated', 'Unique ID', 'Post'}) and 4 missing columns ({'rule_invoked', 'difficulty', 'reasoning', 'decision'}).
This happened while the csv dataset builder was generating data using
hf://datasets/tanmayvasvani/hindi-misinfo-taxonomy-800/hindi_misinfo_800_labeled.csv (at revision 75a43fc2a6b4a5b999d5f61ed112f00b1a072984), ['hf://datasets/tanmayvasvani/hindi-misinfo-taxonomy-800@75a43fc2a6b4a5b999d5f61ed112f00b1a072984/edge_case_log.csv', 'hf://datasets/tanmayvasvani/hindi-misinfo-taxonomy-800@75a43fc2a6b4a5b999d5f61ed112f00b1a072984/hindi_misinfo_800_labeled.csv']
Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1837, in _prepare_split_single
writer.write_table(table)
~~~~~~~~~~~~~~~~~~^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 765, in write_table
self._write_table(pa_table, writer_batch_size=writer_batch_size)
~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 773, in _write_table
pa_table = table_cast(pa_table, self._schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2369, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2297, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
Unique ID: int64
Post: string
Labels Set: string
global_id: string
Category: string
Reason: string
adjudicated: bool
split: string
-- schema metadata --
pandas: '{"index_columns": [{"kind": "range", "name": null, "start": 0, "' + 1195
to
{'global_id': Value('string'), 'difficulty': Value('string'), 'decision': Value('string'), 'reasoning': Value('string'), 'rule_invoked': Value('string')}
because column names don't match
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1369, in compute_config_parquet_and_info_response
parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
~~~~~~~~~~~~~~~~~~~~~~~~~^
builder, max_dataset_size_bytes=max_dataset_size_bytes
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 948, in stream_convert_to_parquet
builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1683, in _prepare_split
for job_id, done, content in self._prepare_split_single(
~~~~~~~~~~~~~~~~~~~~~~~~~~^
gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
):
^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1839, in _prepare_split_single
raise DatasetGenerationCastError.from_cast_error(
...<4 lines>...
)
datasets.exceptions.DatasetGenerationCastError: An error occurred while generating the dataset
All the data files must have the same columns, but at some point there are 7 new columns ({'Reason', 'Category', 'Labels Set', 'split', 'adjudicated', 'Unique ID', 'Post'}) and 4 missing columns ({'rule_invoked', 'difficulty', 'reasoning', 'decision'}).
This happened while the csv dataset builder was generating data using
hf://datasets/tanmayvasvani/hindi-misinfo-taxonomy-800/hindi_misinfo_800_labeled.csv (at revision 75a43fc2a6b4a5b999d5f61ed112f00b1a072984), ['hf://datasets/tanmayvasvani/hindi-misinfo-taxonomy-800@75a43fc2a6b4a5b999d5f61ed112f00b1a072984/edge_case_log.csv', 'hf://datasets/tanmayvasvani/hindi-misinfo-taxonomy-800@75a43fc2a6b4a5b999d5f61ed112f00b1a072984/hindi_misinfo_800_labeled.csv']
Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
global_id string | difficulty string | decision string | reasoning string | rule_invoked string |
|---|---|---|---|---|
HM1626 | True text circulating alongside external false claim | MISS | Factually true biography; omits the surrounding false-death-claim context it traveled with | R3 |
HM0721 | FAB vs MAN | MAN | Real incident (plane stuck under bridge during transport); invented pilot-landing narrative | R2 |
HM1234 | FAB vs OLD | FAB | Verified: no hospitalization occurred at circulation time; event conjured | R2 |
HM0963 | MAN vs OUT | OUT | generic assertion, no specific claim named | R8 |
HM1005 | No fit | OUT | Pure rhetorical insinuation (bioweapon questions); no checkable claim | OUT definition |
HM0898 | PROP vs OUT | OUT | Subjective framing of real protest; no false factual conclusion manufactured | R8 |
HM1303 | MAN vs FAB | FAB | 20 crore jobs statistic has no real-world anchor; invented number | R2 |
HM0679 | MAN vs FAB — rule-changing case | FAB | Video is a staged imposter performance; artifact manufactured for the deception itself — triggered R2 amendment | R2 (amended v1.1) |
HM0425 | MISS vs MAN | MAN | Real textbook page (national nursing textbook); false attribution to Gujarat state syllabus | R2 |
HM0543 | v1.3 migration (PROP retired) | OUT | partisan spin on real Dubey history, no discrete checkable claim | R9 |
HM0527 | v1.3 migration (PROP retired) | MAN | real valved-N95 advisory, invented profit-motive narrative | R2 |
HM0099 | v1.3 migration (PROP retired) | MISS | genuine GDP figures cherry-picked, comparative context omitted | R3/R9 |
HM0617 | v1.3 migration (PROP retired) | OUT | pure endorsement/opinion, nothing checkable | R9 |
HM0619 | v1.3 migration (PROP retired) | MAN | real UPSC results twisted into false Islamic-Studies claim | R2 |
RULING | OLD vs MAN interpretation | — | pure relocation of matching authentic content → OLD; relocation importing new false actors/identities/motives → MAN | R4/R5 |
RULING-2 | FAB vs MAN for fake identities | — | invented identity grafted onto a REAL video/photo → MAN (anchor exists); wholly fabricated persona or staged artifact with no authentic anchor → FAB | R2 |
HM0309 | SAT vs FAB — origin unverifiable | FAB | comedic riff on real Oli-Ayodhya event but satirical origin untraceable within verification cap; conservative ruling as fabricated quote | R6/R2 |
HM0080 | SAT vs FAB — origin unverifiable | FAB | claimed scripted-skit origin unsourceable; labeled by visible mechanism as staged content | R6/R2 |
AUDIT-1 | Anil Upadhyay template (HM0107, HM0367, HM0414, HM0879, HM0445) | MAN | recurring fictional-MLA identity grafted onto real videos; five FAB labels corrected per RULING-2 | R2 |
AUDIT-2 | Drift check | MAN | HM0162: own reason described added false claims; OLD label corrected to MAN | R4 |
AUDIT-3 | Precedent consistency | MAN | HM0677: real women, false same-person claim; FAB corrected to MAN per HM0778 precedent | R2 |
HM0051 | null | null | null | null |
HM0076 | null | null | null | null |
HM1131 | null | null | null | null |
HM1009 | null | null | null | null |
HM0162 | null | null | null | null |
HM1621 | null | null | null | null |
HM0669 | null | null | null | null |
HM1279 | null | null | null | null |
HM0123 | null | null | null | null |
HM0548 | null | null | null | null |
HM0168 | null | null | null | null |
HM0677 | null | null | null | null |
HM0433 | null | null | null | null |
HM1347 | null | null | null | null |
HM0767 | null | null | null | null |
HM0261 | null | null | null | null |
HM1402 | null | null | null | null |
HM1312 | null | null | null | null |
HM1416 | null | null | null | null |
HM0918 | null | null | null | null |
HM0067 | null | null | null | null |
HM0426 | null | null | null | null |
HM1468 | null | null | null | null |
HM0141 | null | null | null | null |
HM0331 | null | null | null | null |
HM0772 | null | null | null | null |
HM0968 | null | null | null | null |
HM0989 | null | null | null | null |
HM0714 | null | null | null | null |
HM1315 | null | null | null | null |
HM0374 | null | null | null | null |
HM0464 | null | null | null | null |
HM1018 | null | null | null | null |
HM1496 | null | null | null | null |
HM0859 | null | null | null | null |
HM1557 | null | null | null | null |
HM0184 | null | null | null | null |
HM1105 | null | null | null | null |
HM1375 | null | null | null | null |
HM1205 | null | null | null | null |
HM0803 | null | null | null | null |
HM1405 | null | null | null | null |
HM0530 | null | null | null | null |
HM1235 | null | null | null | null |
HM0485 | null | null | null | null |
HM1353 | null | null | null | null |
HM1185 | null | null | null | null |
HM0383 | null | null | null | null |
HM0711 | null | null | null | null |
HM1134 | null | null | null | null |
HM1106 | null | null | null | null |
HM0940 | null | null | null | null |
HM0752 | null | null | null | null |
HM0270 | null | null | null | null |
HM1068 | null | null | null | null |
HM1151 | null | null | null | null |
HM0811 | null | null | null | null |
HM0693 | null | null | null | null |
HM0528 | null | null | null | null |
HM1203 | null | null | null | null |
HM1575 | null | null | null | null |
HM0382 | null | null | null | null |
HM0584 | null | null | null | null |
HM0411 | null | null | null | null |
HM0614 | null | null | null | null |
HM0239 | null | null | null | null |
HM1578 | null | null | null | null |
HM0861 | null | null | null | null |
HM0170 | null | null | null | null |
HM1110 | null | null | null | null |
HM1307 | null | null | null | null |
HM0332 | null | null | null | null |
HM0044 | null | null | null | null |
HM0889 | null | null | null | null |
HM1440 | null | null | null | null |
HM1117 | null | null | null | null |
HM1216 | null | null | null | null |
HM1465 | null | null | null | null |
HM0462 | null | null | null | null |
End of preview.
Hindi Misinformation Taxonomy Dataset (HM-800)
A fine-grained re-annotation of 800 Hindi social media posts from the CONSTRAINT 2021 Hindi Hostility Detection dataset, classified by misinformation mechanism rather than binary fake/real labels, using a journalism-informed taxonomy grounded in Wardle & Derakhshan's (2017) information disorder framework.
Why this dataset exists
Existing Hindi misinformation datasets label content as fake or real. This binary framing treats fabricated propaganda the same as satire shared out of context — fundamentally different phenomena requiring different detection strategies. This dataset asks not whether a post misleads, but how.
Taxonomy
| Label | Definition | n |
|---|---|---|
| MAN (Manipulated) | An authentic event, artifact, or statement exists; the post distorts, misattributes, or lies about it | 298 |
| FAB (Fabricated) | The central claim invents an event, statement, or artifact from nothing; includes staged content manufactured for the deception | 251 |
| OUT (Out of scope) | No checkable factual claim (opinion, abuse, insinuation), or the post appears actually true despite its source label | 173 |
| OLD (Recirculated) | Authentic content displaced in time or event-identity, with no other falsification | 57 |
| SAT (Misrepresented satire) | Content created as satire/parody, circulating stripped of that frame; all instances origin-verified | 17 |
| MISS (Missing context) | Everything stated is accurate; the deception is purely subtractive | 4 |
Annotation process
- Base corpus: posts labeled
fakein CONSTRAINT 2021 Hindi (Bhardwaj et al., 2020), deduplicated (1,632 unique), randomly sampled (n=800, seed 42) - Annotator: first author (journalism & mass communication background); category decisions made solely by the annotator; claim verification used web search and AI-assisted search interfaces (the annotation scheme was never shared with these tools)
- Taxonomy development: 150-post exploration set; two-annotator pilot (50 posts); one category (framing propaganda) retired after 0% inter-annotator agreement; guidelines frozen at v1.3 before main annotation
- Inter-annotator agreement: Cohen's κ = 0.615 (raw 73%) on a 100-post overlap independently annotated by a second annotator, computed on pre-adjudication labels
- Adjudication: all 27 disagreements resolved by evidence-based discussion; 18 labels changed; resolution notes preserved in the
Reasonfield - Audit trail: annotation guidelines, edge-case log, and per-post reasoning included
Fields
| Field | Description |
|---|---|
Unique ID |
Original CONSTRAINT 2021 post ID (restarts per split file in source) |
Post |
Hindi post text, unmodified |
Labels Set |
Original CONSTRAINT multi-label annotation |
global_id |
Stable ID assigned in this work (HM0000–HM1631 over the deduplicated fake pool) |
Category |
Taxonomy label (this work's contribution) |
Reason |
Annotator's one-line justification; adjudicated posts carry the resolution note |
adjudicated |
Whether the label was settled through two-annotator adjudication |
split |
train / val / test (600/100/100, stratified by Category, seed 42) |
Known limitations
- Content reflects 2020–21 Indian social media (COVID-19, CAA protests); AI-generated misinformation is unrepresented
- MISS (n=4) and SAT (n=17) are too scarce for reliable per-class model metrics; we recommend excluding them from classification experiments and reporting them descriptively
- OUT posts (21.6%) indicate the source dataset's
fakelabel sometimes marks hostility rather than falsifiability — a dataset-quality finding of this work - Primary annotation was performed by a single domain-expert annotator; the 100-post overlap provides the reliability estimate
- Main annotation occurred across sessions including one extended (~12 hour) session, a deviation from the 90-minute session protocol specified in the guidelines
Source and licensing
Base texts from the CONSTRAINT 2021 Hindi Hostility Detection dataset (Bhardwaj et al., 2020, arXiv:2011.03588); please cite the original work alongside this dataset. Taxonomy annotations released under CC-BY-4.0.
How to load
from datasets import load_dataset
ds = load_dataset("tanmayvasvani/hindi-misinfo-taxonomy-800",
data_files="hindi_misinfo_800_labeled.csv")
Citation
@dataset{hm800_2026,
title = {Hindi Misinformation Taxonomy Dataset (HM-800): Fine-Grained
Journalism-Informed Annotation of Hindi Social Media Misinformation},
author = {Tanmay Vaswani and Roshan Bhatia},
year = {2026},
url = {https://huggingface.co/datasets/[tanmayvasvani]/hindi-misinfo-taxonomy-800}
}
- Downloads last month
- 13