---
dataset_info:
  features:
  - name: sentence1
    dtype: string
  - name: sentence2
    dtype: string
  - name: labels
    dtype: string
  - name: metadata
    dtype: string
  splits:
  - name: train
    num_bytes: 5977000
    num_examples: 11412
  - name: validation
    num_bytes: 782465
    num_examples: 1489
  - name: test
    num_bytes: 743401
    num_examples: 1417
  download_size: 1930694
  dataset_size: 7502866
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
license: mit
source_datasets: []
---

# sentfin

SEntFiN entity sentiment, one native headline/entity judgment per row.

Pinned original CSV: https://github.com/pyRis/SEntFiN (repository MIT license).
Only malformed dictionary quoting is repaired; the original annotation string
and source row ID remain in metadata. Identical normalized headlines stay in
one deterministic 80/10/10 internal partition. These are derived holdouts, not
an official benchmark split. Duplicate pairs are deduplicated; contradictory
judgments for the same headline/entity are excluded and reported.

Original data: [original distribution](https://raw.githubusercontent.com/pyRis/SEntFiN/eba43610fe3cb57a3e5773cd6c037da1f8992ad5/SEntFiN.csv). Repackaged as parquet for [tasksource](https://github.com/sileod/tasksource) by [scripts/repackage_dataset/relevance.py](https://github.com/sileod/tasksource/blob/main/scripts/repackage_dataset/relevance.py#L62).

[Conversion provenance](provenance.json): source revisions, hashes, configuration and exclusions.