abhiprd2000/Bhojpuri-Behavioral-Corpus-8K

PURVA Bhojpuri Sentiment Corpus (v2)

Sentiment labels for native Bhojpuri text, built by having five open-weight LLMs independently judge each sentence and aggregating their disagreement into a consensus label — rather than relying on a single model or a small set of human annotators. This release carries per-sentence labels, 5-judge votes, two consensus methods (majority vote and Dawid–Skene), content hashes for reproducible rehydration, and a 247-item human-labeled gold set for evaluation. Sentence text itself is included only for Bhojpuri Wikipedia rows (CC BY-SA); every other row ships as a hash + source URL — see rehydrate.py.

At a glance

• 90,207 Bhojpuri sentences
• 12 source collections from 11 websites (news, encyclopedic, literary, devotional/commentary, opinion) — kavitakosh and kavitakosh_geet are both kavitakosh.org, see "Known issues"
• 451,002 individual judge labels (5 judges × 90,207 sentences, minus 33 dropped for judge parse failures — see n_agree below)
• 4 sentiment classes: objective, positive, negative, neutral
• 247 sentences with a human gold label (test-split only, pulled from the main 90,207 — not additional text)

Files

File	Rows	What it is
purva_labels.jsonl	90,207	Every sentence: per-judge votes, both consensus labels, split, source metadata. cleaned_text included only for Bhojpuri Wikipedia rows (CC BY-SA).
gold_human.jsonl	247	Human sentiment labels, pulled from three draws: a difficulty-balanced test set, an unanimous-committee-agreement control sample, and a high-disagreement contested sample. Same text rule as above.
rehydrate.py	—	Fetches source_url live and re-derives the sentence text for rows that don't carry cleaned_text, matching by text_sha256. Also in the repo at purva/release/rehydrate.py.

purva_labels.jsonl fields

• id — stable item id (sha1-derived, 16 hex chars)
• text_sha256 — sha256 of the sentence's cleaned text; the join key for rehydrate.py
• source_name, source_url — which of the 12 source collections, and the specific page
• register — news / encyclopedic / literary / verse / commentary / opinion
• text_type — prose or verse
• license_class — redistributable_cc_by_sa (Bhojpuri Wikipedia only) or pointer_only (every other source)
• split — train / dev / test (80/10/10, stratified by register × text_type × source_name; every gold-labeled item is forced into test)
• votes — {judge_name: {label, confidence}} for aya, gemma, llama, mistral, qwen. label is that judge's own 4-class call (subjectivity collapsed with polarity: objective if the judge called it objective, else its polarity). null for the 33 label-judge pairs where the judge's output failed to parse.
• majority_vote — the simple-majority consensus label across the 5 judges
• dawid_skene — the Dawid–Skene EM consensus label (models each judge's reliability; see "Known issues" below for how this compares to majority vote)
• entropy_norm — normalized entropy of the Dawid–Skene posterior (0 = unanimous-equivalent confidence, 1 = maximally uncertain); the project's standard difficulty measure
• n_agree — how many of the 5 judges' own labels match majority_vote (0–5)
• cleaned_text — only present for source_name == "bhojpuri_wikipedia" rows (8,504 of them), under CC BY-SA. Everything else: rehydrate from source_url.

gold_human.jsonl fields

Same id / text_sha256 / source_name / register / text_type / split / cleaned_text-only-for-Wikipedia convention as above, plus:

• human_label — one of objective, positive, negative, neutral, mixed (mixed is a human-only label; no judge or consensus method in this release ever outputs it)
• gold_set — which draw this item came from: gold_test_120 (111 items, difficulty-balanced, verified 100% inside the test split — this is the one to benchmark against), gold_unanimous_91 (91 items, all 5 judges agreed), or gold_contested_44 (44 usable items, high judge disagreement — 45 drawn, 1 excluded as mixed)
• gold_source — the original draw's internal name
• committee_unanimous_label — only present for gold_unanimous_91 items

Only gold_test_120 is guaranteed leak-free against split. gold_unanimous_91 and gold_contested_44 predate the train/dev/test split and land across all three (73 train, 8 dev, 55 test combined) — don't use them as a held-out test set on their own.

Rehydration match rate

rehydrate.py tested against 20 random rows per non-Wikipedia source (seed 42): live-fetches source_url, re-splits and re-cleans the page, and checks for a sentence hashing to text_sha256.

Source	Matched
bhojpurinews	20/20 (100%)
parichaydas_bihari	20/20 (100%)
bhojpuri_blogspot	19/20 (95%)
jogira	18/20 (90%)
sirijan	18/20 (90%)
bhojpurisahityasarita	13/20 (65%)
openbooksonline	12/20 (60%)
kavitakosh_geet	9/20 (45%)
kavitakosh	8/20 (40%)
khabarbhojpuri	0/20 (0%)
pandiji	0/20 (0%)

khabarbhojpuri's 0% is a site-template change: the page's current builder markup no longer has the CSS class the collector originally selected on. pandiji's 0% is a different Blogger template (dynamic/JS-rendered views) whose post body isn't in the static HTML at all. Neither is a sign the underlying text is wrong — both are live-page structure having moved on since the original scrape. A below-100% rate elsewhere is a mix of normal page edits/removals since collection and sentence-boundary differences between the original split and a fresh one.

Splits

Stratified 80/10/10 by (register, text_type, source_name), seed 42:

Split	Sentences
train	72,168
dev	9,019
test	9,020

Licence

This release (labels, votes, consensus annotations, split assignments, hashes, and the bhojpuri_wikipedia text) is CC BY-SA 4.0.

That covers the annotations on every row and the text of the 8,504 Bhojpuri Wikipedia rows. It does not relicense the other 81,703 rows' original text — those sites were collected under a pointer-only basis (see SOURCES.md), so this release carries only source_url + text_sha256 for them. Run rehydrate.py against the live site to recover that text yourself.

Label definitions

Two-stage scheme (full annotation guidelines in ANNOTATION_GUIDELINES.md):

• objective — reports a fact, event, or description with no evaluative stance. Typical of news and encyclopedic register.
• positive — subjective and favorable: appreciative, hopeful, celebratory.
• negative — subjective and unfavorable: critical, sorrowful, angry, disapproving.
• neutral — subjective (it is a stance, not a bare fact) but mild, hedged, or otherwise without clear positive/negative charge.
• (human-gold only) mixed — clearly both positive and negative, neither dominant. No judge, majority-vote, or Dawid–Skene output in this release ever produces mixed — it only appears in gold_human.jsonl's human_label.

Known issues

• Neutral is rare and the hardest class. Only 733/90,207 (0.8%) of majority-vote labels are neutral. On the 108-item scored gold set, per-class recall on neutral is 11.8% (majority vote) / 23.5% (Dawid–Skene) — far below the other three classes. Don't expect strong neutral detection from anything trained on this label space without dedicated work on that class.
• Majority vote beats Dawid–Skene on the held-out human gold set, despite DS being the more sophisticated aggregator: accuracy 0.657 vs 0.574 (+0.083, 95% CI [+0.018, +0.157]); macro-F1 0.531 vs 0.494 (+0.037, CI crosses zero). The two consensus methods disagree on 22/108 gold items. Both are included in purva_labels.jsonl; pick majority vote unless you have a specific reason to prefer DS.
• Verse is harder than prose. Accuracy on the verse register is 0.50–0.39 across models, versus 0.81–0.83 on news and encyclopedic. Verse sentences here are typically single lines of poetry, not full sentences, which may be part of why.
• kavitakosh_geet is a collection label, not a separate site. These 12,272 rows come from kavitakosh.org, from a second run of the same collector after a fix to how it walks category pages. The source_name label kept an old name. A live check of 100 random rows found every page still online, and 52% of sentences re-matched exactly (42% for the other kavitakosh rows), which is normal for verse, where line splits vary.

Version history

• v1-legacy (tag v1-legacy): a 68,822-row, artificially 1:1:1 class-balanced corpus with a substantial machine-translated/anchored- translation component (English terms bracketed inline with Bhojpuri). Used to train the v1 model (XLM-RoBERTa, 97.6% accuracy on its own 501-item balanced/translated test set).
• v2 (this release): replaces it with labels over native, non-translated text (not included directly — see "Licence") and a naturally imbalanced label distribution, scored against a human-labeled gold set rather than a held-out slice of the same synthetic/translated data. v1 and v2 numbers are not comparable — different text, different label source, different test set.

Citation

@misc{prasad2026purva,
  author = {Prasad, Abhimanyu},
  title = {PURVA: Bhojpuri Sentiment Corpus},
  year = {2026},
  howpublished = {\url{https://github.com/shodhx/purva}},
}