ctu-aic/nli_it_collection

Dataset Card for Natural Language Inference Instruction Tuning Collection

This dataset is a collection of various NLI datasets in Czech and English, transformed into an instruction tuning format based on the FLAN approach.

Dataset Details

Dataset Description

This dataset is a collection of English and Czech NLI datasets. Its primary purpose is instruction tuning (supervised fine tuning) of decoder LLMs. The used datasets were converted using a FLAN-like templates.

• Curated by: Artificial Intelligence Center, FEE, CTU in Prague
• Language(s) (NLP): Czech (cs, ces), English (en)
• License: cc-by-nc-sa-4.0

Dataset Sources

The dataset consists of the following datasets:

English 🇺🇸 🇬🇧

• FEVER - FEVER transformed for NLI
• AVeriTeC - train and development gold splits with concatenated question-answer pairs as the evidence
• SNLI
• ANLI

Czech 🇨🇿

• CsFEVER-NLI - FEVER translated to Czech using Deepl translator
• CtkFACTS-NLI - Original Czech NLI dataset
• SNLI_CS - SNLI translated to Czech using Google translator
• ANLI_CS - ANLI translated to Czech

Uses

The dataset is intended for simillar usage as the original FLAN dataset. Its main purpose is instruction tuning (supervised fine tuning) of decoder LLMs on NLI task.

Direct Use

For direct usage there is column conversations, which can be directly used for training using Transformers and Transformers-related libraries.

Out-of-Scope Use

This collection is not directly intended for training decoder models, however, it can be transformed for this purposes as well.

Dataset Structure

The datasets consists of two language subsets - cs (Czech), en (English) Each of the subsets contains following columns:

• id (str) - identifier, unique only in the dataset of origin
• premise (str) - premise (NLI), evidence (fact-checking)
• hypothesis (str) - hypothesis (NLI), claim (fact-checking)
• answer (str) - correct answer to the NLI/fact-checking question
• dataset (str) - the original datasets, which is source of the data
• translated (bool) - true if the data point is translated from other language
• input (str) - input created using FLAN-like template using premise, hypothesis and answer
• output (str) - expected output created according to the randomly chosed FLAN-like template.
• conversation (List[Dict[str, str]]) - Hugging Face Transformers-compatible conversation style format, composed from input and output, which can be directly used for instruction tuning (LLM instruction template can be directly applied)

Dataset Creation

Curation Rationale

The creation was motivated by the nonexistence of any simillar collection with FLAN-like instructions for the Czech language.

Source Data

For the original data sources please refer to the original datasets.

Bias, Risks, and Limitations

This collection contains the biases, risks and limitations of the underlying datasets. Moreover, other limitation is that the variety of prompt templates is limited to 10 per dataset.

Recommendations

Users should be made aware of the risks, biases and limitations of the dataset.

Citation [TBD]

BibTeX:

@mastersthesis{mlynar2025llmadapt,
  author  = {Tomáš Mlynář},
  title   = {Compute-constrained LLM adaptation to Czech language},
  school  = {Czech Technical University in Prague},
  year    = {2025},
  type    = {Master's thesis},
  month   = {6},
  note    = {Supervisor: Ing. Herbert Ullrich},
  url     = {http://hdl.handle.net/10467/123587}
}

Dataset Card Contact

If there is any problem or question, please use the dataset discussion here on Huggingface.