README.md

metadata

dataset_info:
  features:
    - name: hyponym
      dtype: string
    - name: hypernym
      dtype: string
    - name: definition
      dtype: string
  splits:
    - name: train
      num_bytes: 4974231
      num_examples: 44772
    - name: test
      num_bytes: 5422
      num_examples: 49
  download_size: 3485218
  dataset_size: 4979653
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*

Dataset card for WordNet-TaxoLLaMA

TaxoLLaMA is a model capable of solving Lexical Semantics task with SoTA metrics. The model was fine-tuned on instructive dataset WordNet-TaxoLLaMA. It consists of hypernym-hyponym pairs sampled from WordNet 3.0. As well, it contains definitions, that were used during training to help model disambiguate senses.

Input Format

The TaxoLLaMA model was trained to use the following format :

<s>[INST] <<SYS>> You are a helpfull assistant. List all the possible words divided with a coma. Your answer should not include anything except the words divided by a coma<</SYS>>
hyponym: tiger (large feline of forests in most of Asia having a tawny coat with black stripes)| hypernyms: [/INST]

We recommend you to follow this format, however you are free to change it to suite your task!

Citation

If you find TaxoLLaMA or WordNet-TaxoLLaMA is useful in your work, please cite it with:

@misc{moskvoretskii2024taxollama,
      title={TaxoLLaMA: WordNet-based Model for Solving Multiple Lexical Sematic Tasks}, 
      author={Viktor Moskvoretskii and Ekaterina Neminova and Alina Lobanova and Alexander Panchenko and Irina Nikishina},
      year={2024},
      eprint={2403.09207},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}