IndicIFBench

IndicIFBench is a translation of IFBench, a benchmark for verifiable instruction following, into 14 Indian languages. Each prompt combines a task with one or more instructions whose compliance can be checked automatically (for example "include the keyword river twice" or "answer in exactly 3 paragraphs"). The instructions, and the parameters the checkers use, are translated together so that every prompt stays verifiable in its language. Every prompt was reviewed by a native speaker.

3,822 prompts. Prompts that native speakers marked as incorrect without a usable correction were removed, so the number of prompts differs per language; 190 prompts are present in all 14 languages.

Code	Language	Script	Prompts	Verified	Corrected
as	Assamese	Bengali-Assamese	280	271	9
bn	Bengali	Bengali	279	277	2
gu	Gujarati	Gujarati	281	281	0
hi	Hindi	Devanagari	282	260	22
kn	Kannada	Kannada	276	276	0
ml	Malayalam	Malayalam	275	275	0
mr	Marathi	Devanagari	277	275	2
ne	Nepali	Devanagari	258	258	0
or	Odia	Odia	274	208	66
pa	Punjabi	Gurmukhi	281	276	5
sa	Sanskrit	Devanagari	275	271	4
ta	Tamil	Tamil	271	268	3
te	Telugu	Telugu	246	243	3
ur	Urdu	Perso-Arabic	267	267	0

Loading

from datasets import load_dataset

ds = load_dataset("ai4bharat/IndicIFBench", "hi", split="train")

There is one configuration per language code. To compare languages on the same prompts, keep the keys present in every language you compare.

Data fields

The schema follows IFBench:

Field	Description
key	Row identifier, the same as in English IFBench, so the same prompt can be compared across languages.
prompt	The translated prompt.
instruction_id_list	The IFBench instruction IDs that apply to the prompt.
kwargs	One dictionary of checker parameters per instruction, in the same order. Keywords, answer options and texts to repeat or overlap with are translated and match the prompt exactly.
verification	verified: a native speaker confirmed the translation; corrected: changed according to the native speaker's correction.

Two additions to the IFBench format:

• kwargs[i]._lang: translated material the checker needs for this language. fixed holds translated fixed strings (template labels such as "My Answer:", CSV column names, the sentence to repeat, the answer options' "or", the European capitals) and names holds the transliterated person names of count:person_names.
• response_language:xx: a few prompts ask for the answer in another language (they translate a passage into Russian, or rewrite it in modern English). These rows carry an extra scored instruction, response_language:ru (keys 52, 181) or response_language:en (keys 117, 206, 208, 240, 248), with empty parameters, which checks the answer's language.

How it was built

1. Selection. 282 of the 300 English IFBench prompts. 18 were left out because their instructions cannot carry over to Indic languages: title case and sentences starting with A to Z (Indic scripts have no letter case or A to Z), words in alphabetical order looping from Z back to A, starting with a verb, and an English alphabetical-order task.
2. Translation with Gemini. Keywords, answer options, person names, texts to repeat or overlap with, and fixed strings were translated once and synchronised between the prompt and kwargs. Numbers stay in ASCII digits. Code, formulas, URLs and established acronyms stay as in the source.
3. Adaptation. Instructions on letters, vowels, consonant clusters and syllables state the unit they count in the target script. Prompts that ask to translate, rewrite or check a given text keep that text in its original language (keys 52, 115, 117, 181, 206, 208, 240, 243, 248).
4. Automatic quality control. Every translation was reviewed by an LLM judge, flagged rows were repaired by an LLM corrector, and automatic consistency checks were run on all rows (every parameter appears in its prompt, keywords and names as whole words, keyword counts match the English, no letters from another script, no non-ASCII digits).
5. Human verification. Native speakers compared every translated prompt with the English source and checked that the checker parameters match the prompt. Their corrections were applied, with keywords changed in the prompt and the parameters together (116 prompts, verification: corrected); 3,706 prompts were confirmed unchanged (verified). Prompts marked incorrect without a usable correction were removed.

Evaluation

The data uses IFBench's instruction IDs and parameter format. Many original IFBench checkers assume English text (word lists, sentence splitting, letters, keyword boundaries), so scoring Indic responses requires language-aware checkers that read _lang and apply each script's rules.

License

ODC-BY-1.0, the license of the IFBench data, which this dataset translates. It is intended for research and educational use in accordance with Ai2's Responsible Use Guidelines.

Citation

Please cite IFBench, which this dataset is based on:

@misc{pyatkin2025generalizing,
   title={Generalizing Verifiable Instruction Following},
   author={Valentina Pyatkin and Saumya Malik and Victoria Graf and Hamish Ivison and Shengyi Huang and Pradeep Dasigi and Nathan Lambert and Hannaneh Hajishirzi},
   journal={Advances in Neural Information Processing Systems},
   volume={38},
   year={2025}
}