MattiaL/tapir-cleaned-67k

Dataset Card for Tapir-Cleaned

This is a revised version of the DAISLab dataset of IFTTT rules, which has been thoroughly cleaned, scored, and adjusted for the purpose of instruction-tuning.

Tapir Dataset Summary

Tapir is a subset of the larger DAISLab dataset, which comprises 242,480 recipes extracted from the IFTTT platform.

After a thorough cleaning process that involved the removal of redundant and inconsistent recipes, the refined dataset was condensed to include 67,697 high-quality recipes.

This curated set of instruction data is particularly useful for conducting instruction-tuning exercises for language models, allowing them to more accurately follow instructions and achieve superior performance.

The last version of Tapir includes a correlation score that helps to identify the most appropriate description-rule pairs for instruction tuning. Description-rule pairs with a score greater than 0.75 are deemed good enough and are prioritized for further analysis and tuning.

Supported Tasks and Leaderboards

The Tapir dataset designed for instruction training pretrained language models

Languages

The data in Tapir are mainly in English (BCP-47 en).

Dataset Structure

Data Instances

{
    "instruction":"From the description of a rule: identify the 'trigger', identify the 'action', write a IF 'trigger' THEN 'action' rule.",
    "input":"If it's raining outside, you'll want some nice warm colors inside!",
    "output":"IF Weather Underground Current condition changes to THEN LIFX Change color of lights",
    "score":"0.788197",
    "text": "Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n\n### Instruction:\nFrom the description of a rule: identify the 'trigger', identify the 'action', write a IF 'trigger' THEN 'action' rule.\n\n### Input:\nIf it's raining outside, you'll want some nice warm colors inside!\n\n### Response:\nIF Weather Underground Current condition changes to THEN LIFX Change color of lights",
}

Data Fields

The data fields are as follows:

• instruction: describes the task the model should perform.
• input: context or input for the task. Each of the 67K input is unique.
• output: the answer taken from the original Tapir Dataset formatted as an IFTTT recipe.
• score: the correlation score obtained via BertForNextSentencePrediction
• text: the instruction, input and output formatted with the prompt template used by the authors of Alpaca for fine-tuning their models.

Data Splits

	train
tapir	67697

Licensing Information

The dataset is available under the Creative Commons NonCommercial (CC BY-NC 4.0).

Citation Information

@misc{tapir,
  author = {Mattia Limone, Gaetano Cimino, Annunziata Elefante},
  title = {TAPIR: Trigger Action Platform for Information Retrieval},
  year = {2023},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/MattiaLimone/ifttt_recommendation_system}},
}