idrock/ifeval-uzbek

IFEval Uzbek — Latin Script

A fully translated version of google/IFEval into Uzbek Latin script, preserving every instruction constraint, format requirement, and structural element of the original prompts.

Dataset Summary

Language	Uzbek (Latin script)
Source dataset	google/IFEval
Split	train
Rows	541
Script	Latin only (no Cyrillic)
Translation model	mercury-2 (Inception Labs)
Manually reviewed & corrected	144 / 581 rows

Motivation

IFEval is a benchmark for evaluating instruction-following ability in LLMs. This dataset makes it possible to evaluate Uzbek-language models on the same structured benchmark without relying on English prompts.

Dataset Structure

Columns

Column	Type	Description
key	int	Original IFEval row key
prompt	string	Original English prompt
prompt_uz	string	Full Uzbek Latin translation of the prompt
instruction_id_list	list	IFEval instruction type IDs
kwargs	list	Instruction parameters (counts, keywords, phrases, etc.)
translation_meta	dict	Translation provenance (model, attempt, timestamps)

Example

from datasets import load_dataset

ds = load_dataset("Dovud-Asadov/ifeval-uzbek-prompt-latin")
print(ds["train"][0]["prompt_uz"])

Translation Methodology

1. Automatic translation — all 581 prompts were translated using mercury-2 (Inception Labs) with a strict system prompt enforcing:

   • Full semantic fidelity (no omissions, no paraphrasing)
   • Uzbek Latin script only (no Cyrillic)
   • Preservation of literal tokens: URLs, placeholders [name], code fences, angle markers, *** dividers, quoted strings, and forbidden/required English keyword literals

2. Manual review & correction — every translation was reviewed individually. 144 rows were corrected for issues including:

   • Model executing the prompt instead of translating it (generated emails, essays, poems, etc.)
   • Model refusing to translate (returning apologies)
   • Cyrillic characters mixed into Latin output
   • Wrong language output (Kazakh, Turkish, etc.)
   • Missing instruction sections or garbled text
   • Critical mistranslations that reversed the meaning of constraints

Instruction Type Distribution (top 10)

Instruction type	Count
punctuation:no_comma	73
length_constraints:number_words	59
length_constraints:number_sentences	53
keywords:forbidden_words	52
combination:repeat_prompt	50
detectable_format:number_highlighted_sections	49
keywords:frequency	43
startend:quotation	42
keywords:existence	41
change_case:english_lowercase	40

Translation Notes

• Preserved as-is: URLs, placeholders like [name], code/markdown tokens, angle-bracket markers, *** dividers, specific English keyword strings that the model must include or exclude in its response (e.g. farmer, economy), required English end-phrases (e.g. "Is there anything else I can help with?"), and quoted source texts meant to be rewritten.
• Translated: all natural-language instruction sentences, topic descriptions, and format directives.
• The translation_meta.manually_corrected field is true for rows that received a human correction.

Citation

If you use this dataset, please also cite the original IFEval paper:

@article{zhou2023instruction,
  title={Instruction-Following Evaluation for Large Language Models},
  author={Zhou, Jeffrey and Lu, Tianhao and Mishra, Swaroop and Brahma, Siddhartha and Basu, Sujoy and Luan, Yi and Zhou, Denny and Hou, Le},
  journal={arXiv preprint arXiv:2311.07911},
  year={2023}
}

License

Apache 2.0 — same as the original IFEval dataset.