tabiya/esco_skills_test

Original Repository: Skill-Extraction-benchmark

Download

To download the dataset create an access read token.

Run the following script to download the model using the datasets library:

from datasets import load_dataset
HF_TOKEN = "hf_..."
skills = load_dataset('tabiya/esco_skills_test', token=HF_TOKEN)

Optional

Convert to pandas dataframe:

import pandas as pd
df_skillspan = skills['skillspan'].to_pandas()
df_techwolf = skills['techwolf'].to_pandas()

Description

IMPORTANT: Please report MAP@1 and MAP@5 for comparative evaluation!

This dataset is used for evaluating retrieval methods from the ESCO skills taxonomy. Two subsets:

1. skillspan: 2406 skills contained in 920 sentences extracted from job postings
2. techwolf: Skill phrases mapped to ESCO

The original datasets (skillspan-00000-of-00001.parquet and techwolf-00000-of-00001.parquet) contain the following columns

• sentence: text from the Skillspan dataset
• span: the relevant skill to be mapped
• subspan: possible (non necessarily continuous) subtext from the skill span referring to the skill of interest
• label: esco skill label of the leaf corresponding to the skill of interest

The span and subspan columns are empty in the techwolf database.

The datasets skillspan_with_id.parquet and techwolf_with_id.parquet are artifacts obtained from the previous corresponding datasets by adding he following two columns:

• ID: unique identifier for the row in the original dataset
• UUID: unique identifier of the skill label, obtained by mapping the label string to the one found in the Tabiya ESCO 1.1.1 version.

The skillspan test dataset contains 981 rows in which the skill label is unknown (UNK) and therefore the corresponding UUID is None.

Finally, the processed_skill_test_set_with_id.parquet file consists of the merged two previous datasets, without the rows with UNK value in their skill label. In addition to the previous columns, a synthetic_query column has been added to allow for evaluation of the skill vector search component of Compass. This consists of a re-elaboration of the sentence entry by asking a Large Language Model to write the answer to the question "What are your skills and expertise?" of someone having the particular skill named in sentence.