ai4bharat/human-eval

Airavata HumanEval Prompts

This benchmark contains a set of prompts written by real-users to evaluate LLMs on real-world tasks and test it for different abilities. We collect prompts for 5 abilities listed below:

• Long: Ability to generate long-form text like writing essays, speeches, reports, etc.
• Fact-Ops: Ability to give factual opinions and explanations like seeking recommendations, seeking advice, opinions, explanations, etc.
• Content: Ability to make content accessible like summarizations, layman explanations, etc
• Lang-Creativity: Ability to be creative in language like finding anagrams, rhyming words, vocabulary enhancement, etc
• Culture: Ability to answer questions related to Indian Culture.

For each ability we define a list of intents and domains which are provided to the users along with detailed instructions about what prompts are expected.

We recommend the readers to check out our official blog post for more details.

Citation

@misc{airavata2024,
  title = {Introducing Airavata: Hindi Instruction-tuned LLM},
  url = {https://ai4bharat.github.io/airavata},
  author = {Jay Gala and Thanmay Jayakumar and Jaavid Aktar Husain and Aswanth Kumar and Mohammed Safi Ur Rahman Khan and Diptesh Kanojia and Ratish Puduppully and Mitesh Khapra and Raj Dabre and Rudra Murthy and Anoop Kunchukuttan},
  month = {January},
  year = {2024}
}