MaelisResearch/odia-eval-benchmark

Odia Eval Benchmark

Dataset Summary

odia-eval-benchmark is a comprehensive, curated evaluation benchmark for Odia (Oriya) natural language understanding and generation. It consolidates 34 publicly available Odia datasets into a single, normalized format covering 7 task families and 121,947 evaluation rows.

This benchmark was built from authoritative sources with three major improvements:

1. Curated sources - 8 datasets were re-pulled from their original/authoritative sources to fix broken links and corrupt data
2. Machine-translated choices - English multiple-choice options were translated to Odia with NLLB-200 (1.3B), enabling fully Odia-native MCQ evaluation
3. Contamination control - gold answers and references were screened against large Odia pretraining corpora; any row whose gold content a pretrained model may have memorized was excluded, so evaluation reflects true Odia capability rather than training-data recall

Design decisions:

• Fixed source URLs and re-pulled 8 datasets (SQuAD-v2, SQuAD-shifts, IndicQA, winogrande, truthfulqa, xsum, CNN-DailyMail, BBC XSum) from their canonical sources
• Dropped 3,865 SQuAD-v2 empty-gold rows and 401 IndicQA rows with corrupt answers
• Dropped 110 MMLU rows with corrupt answer mappings
• Translated 91,713 English MC options to Odia (NLLB-200, validated)
• Excluded 1,861 rows whose gold content overlaps pretraining data (contamination)
• Recovered 10,642 missing answer indexes (hellaswag + winogrande) by matching gold references to choices
• Regenerated globally unique row ids
• Added script column (Odia / Latin) for script-agnostic evaluation

Languages

• Language: Odia (or, ISO 639-3: ory)
• Script: Odia (Odia-script); 9 rows use Latin-script math/code formulas (kept verbatim)

Dataset Structure

Data Fields

Field	Type	Description
dataset_name	string	Source dataset identifier
task	string	Task family: multiple_choice, qa, generation, math, classification, ner, translation
language	string	ory
split	string	test or validation
quality_tier	string	native, translated, or professional
source_url	string	Canonical source URL
license	string	cc-by-4.0 or mit
prompt	string	English prompt (if available)
prompt_odia	string	Odia prompt
reference / reference_odia	string	Reference / gold answer (EN / OD)
context_en / context_odia	string	Context passage (EN / OD)
choices / choices_odia	list[string]	MC options (EN / OD)
answer	string	Gold answer text
answer_index	int	Index of gold choice (MC)
answer_letter	string	Letter of gold choice (MC)
answers / answer_start	list	Extractive QA gold spans
answer_type	string	index, text, span, or bool
is_bool	bool	Yes/no question flag
num_options	int	Number of MC options
tokens / ner_tags / ner_scheme	list / string	NER tokens and BIO tags
label / label_index / label_names	string / int / list	Classification labels
source_text / target_text	string	Translation source / target
domain	string	Content domain
subset / question_type	string	Source subset / question type
id	string	Globally unique row id ({dataset}/{index})
script	string	Odia or Latin

Task Distribution

Task	Rows	Share
Multiple choice	64,031	52.5%
QA (extractive / boolean / open)	26,997	22.1%
Generation	22,081	18.1%
Classification	4,043	3.3%
Math	2,638	2.2%
NER	1,147	0.9%
Translation	1,010	0.8%

Splits

Split	Rows
test	105,287
validation	16,660

Quality Tiers

Tier	Rows	Meaning
professional	47,908	Professionally translated by Sarvam AI
native	47,075	Natively Odia (collected in Odia)
translated	26,964	Community-translated (tripathysagar)

Included Datasets

Dataset	Task	Rows	Split	License
bhram_il	QA	2,087	test	cc-by-4.0
flores_plus	Translation	1,010	test	cc-by-4.0
indic_copa	MC	500	test	cc-by-4.0
indic_glue_csqa	QA	1,975	test	cc-by-4.0
indic_glue_ner	NER	153	test	cc-by-4.0
indic_glue_wstp	MC	501	test	cc-by-4.0
indic_headline_gen	Generation	6,485	test	cc-by-4.0
indic_qa	QA	1,276	test	cc-by-4.0
indic_quest_odia	QA	199	test	cc-by-4.0
indic_question_generation	Generation	9,973	test	cc-by-4.0
indic_sentence_summarization	Generation	5,623	test	cc-by-4.0
indic_squad_odia	QA	7,736	test	cc-by-4.0
indic_xparaphrase	Classification	1,995	test	cc-by-4.0
milu	MC	4,520	test	cc-by-4.0
naamapadam_odia_ner	NER	994	test	cc-by-4.0
odia_arc	MC	1,170	test	cc-by-4.0
odia_gsm8k	Math	1,319	test	cc-by-4.0
odia_hellaswag	MC	19,825	test	cc-by-4.0
odia_news_classification	Classification	2,048	test	cc-by-4.0
odia_truthfulqa	QA	807	validation	cc-by-4.0
odia_truthfulqa_mc	MC	809	validation	cc-by-4.0
odia_winogrande	MC	3,034	test	cc-by-4.0
sarvam_arc_challenge	MC	1,150	test	mit
sarvam_arc_challenge_val	MC	294	validation	mit
sarvam_boolq	QA	9,427	test	mit
sarvam_boolq_val	QA	3,270	validation	mit
sarvam_gsm8k_indic	Math	1,319	test	mit
sarvam_indivibe_chat	QA	100	test	mit
sarvam_indivibe_code	QA	40	test	mit
sarvam_indivibe_math	QA	40	test	mit
sarvam_indivibe_stem	QA	40	test	mit
sarvam_mmlu_indic	MC	14,003	test	mit
sarvam_mmlu_indic_val	MC	281	validation	mit
sarvam_triviaqa	MC	17,944	test	mit

Dataset Creation

Why this dataset?

Odia (ISO 639-3: ory) is spoken by ~38 million people, yet it is severely underrepresented in NLP evaluation. Existing benchmarks are scattered across repositories with inconsistent formats, broken links, and English-only prompts for multiple-choice tasks. This benchmark unifies them.

Contamination Control

Gold content (answers, references, target translations, gold choices) was screened for substring containment against a large-scale Odia web-text corpus (12.4M+ rows). A row was excluded if:

• its gold text matched a corpus row exactly (>= 40 chars), or
• every 64-char window of its gold text appeared in a single corpus shard (i.e., the gold text is genuinely contained in the corpus)

Rows whose input (prompt/context) matched the corpus but whose gold did not were kept - input passages being in the pretraining corpus is expected and does not leak the answer. This distinction is why only 1,861 of 15,236 raw matches were excluded. Details:

• Excluded: 1,861 rows (gold matches), dominated by headline generation (652), question generation (580), and sentence summarization (425)
• Kept: 13,375 rows with benign input-only overlap

Translation

English MC options (91,713 unique strings, 100,816 occurrences) were translated to Odia with NLLB-200 (distilled 1.3B, int8) using CTranslate2 with length-band batching. A validation gate checked for:

• empty output (253 -> fell back to source)
• untranslated output (1946 math/chem formulas -> kept verbatim)
• blown-up short strings (103 -> fell back to source)
• sibling choice collisions (19 -> fell back to source)

Math and chemistry formulas (32°, H₂ + O₂, NaCl) are kept in their original form - they are language-independent.

Quality Gates

• 0 rows with out-of-bounds or missing answer index (10,642 recovered via gold-reference matching)
• 0 rows with no Odia content at all
• 100% of MCQ rows carry a valid answer_index
• Globally unique row ids

Usage

With Hugging Face datasets

from datasets import load_dataset

ds = load_dataset("MaelisResearch/odia-eval-benchmark", split="train")

# MCQ rows with fully Odia options
mc = ds.filter(lambda x: x["task"] == "multiple_choice")
print(mc[0])

Evaluation

The benchmark is designed to be used with an Odia-aware evaluation harness. A reference harness (eval_harness.py) is maintained in the associated benchmark repository and scores:

• Multiple choice - accuracy on answer_index
• QA - span F1 (SQuAD-style) for extractive; exact-match for boolean
• Math - exact match on the final number
• Classification - accuracy on label_index
• NER - token-level BIO F1
• Translation - chrF / BLEU
• Generation - reference-free (optional LLM judge)

Considerations and Limitations

• Contamination is controlled, not eliminated - the screen covers a specific Odia web-text corpus; overlap with other corpora is possible
• Translated MC options may differ slightly from native phrasing - models that rely on English-style semantics may see a small distribution shift
• Math/chem formulas are intentionally untranslated (language-independent)
• IndiVibe rows (220) are judge-scored conversation/chat items with no gold answer - they are included for reference-style evaluation
• quality_tier lets you filter by provenance: use professional for the highest-quality subsets, native for natively collected data

License

Each row carries its source license (cc-by-4.0 or mit); consumers must respect the license of each constituent dataset. When redistributing, retain the license and source_url fields.

Citation

@misc{maelis-odia-eval,
  title = {Odia Eval Benchmark},
  author = {Maelis Research},
  year = {2026},
  howpublished = {https://huggingface.co/datasets/MaelisResearch/odia-eval-benchmark}
}

Contact

Maelis Research - publishing high-quality Indic language resources for open NLP research.

With this, we begin our journey towards revolutionising AI for the Odia language. Jai Jagannath!