---
dataset_info:
  features:
  - name: sentence1
    dtype: string
  - name: sentence2
    dtype: string
  - name: labels
    dtype: string
  - name: metadata
    dtype: string
  splits:
  - name: train
    num_bytes: 275433385
    num_examples: 10006
  - name: validation
    num_bytes: 34474382
    num_examples: 1231
  - name: test
    num_bytes: 32735387
    num_examples: 1215
  download_size: 126382724
  dataset_size: 342643154
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
license: other
source_datasets: []
---

# evidence-inference

Evidence Inference v2: full article + intervention/comparator/outcome → effect direction.

Join annotations by PromptID AND PMCID, never by article alone. Use verified
labels, exclude tied votes and upstream known-bad prompts. Evidence spans in
metadata refer only to the same prompt and winning label. Full article text
stays in the input so isolated snippets need not identify treatment arms.
Native article train/validation/test splits are preserved. The upstream code
and annotations have MIT terms; article-text rights vary and are not relabeled
MIT. Mirror text/data licensing remains unspecified.

Original data: [original distribution](https://evidence-inference.ebm-nlp.com/v2.0.tar.gz). Repackaged as parquet for [tasksource](https://github.com/sileod/tasksource) by [scripts/repackage_dataset/specialized_text.py](https://github.com/sileod/tasksource/blob/main/scripts/repackage_dataset/specialized_text.py#L90).

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