prajaktakini/folio

Overview

This is a filtered version of the FOLIO (First-Order Logic Inference) dataset, containing only examples where the gold FOL premises and conclusion can be correctly evaluated by Prover9, a first-order logic theorem prover. This filtered dataset serves as a "SKYLINE" baseline, ensuring that all examples have valid, parseable FOL that Prover9 can process correctly.

Dataset Creation Process

The filtered dataset was created by running the fol-gold-to-prover9 experiment, which:

1. Takes the gold FOL premises and conclusion from each example
2. Evaluates them using Prover9
3. Compares Prover9's output with the gold answer (True/False/Uncertain)
4. Retains only examples where Prover9's evaluation matches the gold answer

Source Datasets

The filtered dataset combines examples from two sources:

• Validation Split: Loaded from benlipkin/folio (validation split)

  • This dataset has approximately 93% accuracy when evaluated with Prover9
  • All correct examples from this validation split are included

• Train Split: Loaded from yale-nlp/FOLIO (train split)

  • Up to 200 correct examples are included (or fewer if not enough correct examples are found)
  • Processing stops early once 200 correct examples are collected

Filtering Criteria

An example is included in the filtered dataset if:

1. The gold FOL premises and conclusion can be successfully parsed and evaluated by Prover9
2. Prover9's output (True/False/Uncertain) matches the gold answer exactly
3. No errors occur during FOL parsing or Prover9 evaluation

Examples are excluded if:

• Prover9 produces an error during evaluation
• Prover9's output does not match the gold answer
• The FOL cannot be parsed or reformatted correctly

Dataset Structure

The filtered dataset consists of two splits:

Validation Split (validation.json)

• Source: benlipkin/folio validation split
• Content: All examples from the source validation split where fol-gold-to-prover9 produces the correct answer
• Purpose: Used for validation/testing in experiments

Train Split (train.json)

• Source: yale-nlp/FOLIO train split
• Content: Up to 200 correct examples (or fewer if not enough are found)
• Purpose: Used for training/few-shot examples in experiments

Data Format

Each example in the dataset follows the standard FOLIO format:

{
  "id": "example_id",
  "premises": ["premise 1 in natural language", "premise 2 in natural language", ...],
  "conclusion": "conclusion statement in natural language",
  "premises-FOL": ["premise1_fol", "premise2_fol", ...],
  "conclusion-FOL": "conclusion_fol",
  "label": "True" | "False" | "Uncertain",
  ...
}

Statistics

The exact number of examples in each split depends on the filtering results:

• Validation: All correct examples from benlipkin/folio validation split
• Train: Up to 200 correct examples from yale-nlp/FOLIO train split