Tabfact Eval

tabfact-eval

Chain-of-Table: Evolving Tables in the Reasoning Chain for Table Understanding — Wang et al. (2024) (arXiv:2401.04398, 2024)

What this evaluates

Tests a model's ability to verify the truthfulness of a factual statement given a table. It probes structured reasoning capabilities by requiring the model to cross-reference table contents with a claim and output a binary label.

Datasets

• TabFact — total ?; splits: test (-1)

Metrics

• binary classification accuracy (primary) — range: [0, 1]
  • Percentage of correctly predicted truth labels (True/False) compared to the ground truth labels.

Input / output format

Input: A tabular dataset and a factual statement.

Output: A binary label indicating whether the statement is True or False based on the table.

Scoring recipe

def score(pred, gold):
    return 1.0 if pred == gold else 0.0
accuracy = sum(score(p, g) for p, g in zip(predictions, golds)) / len(golds)

Common pitfalls

• Accuracy is computed over the official test split; using training or validation splits inflates scores.
• The binary classification task requires strict adherence to the True/False labels without generating free-form explanations.

Evidence (verbatim from paper)

TabFact, on the other hand, is a table-based binary fact verification benchmark. The task is to ascertain the truthfulness of a given statement based on the table. ... and for TabFact, we employ the binary classification accuracy.

Citation

@misc{wang2024chainoftable,
  title={Chain-of-Table: Evolving Tables in the Reasoning Chain for Table Understanding},
  author={Wang et al. (2024)},
  year={2024},
  note={arXiv:2401.04398}
}

• arXiv: 2401.04398