Tovo Consensus Eval

tovo-consensus-eval

ToVo: Toxicity Taxonomy via Voting — Luong et al. (2024) (arXiv:2406.14835, 2024)

What this evaluates

This evaluation probes a model's ability to classify text content according to a user-defined toxicity taxonomy. It measures how closely the model's predictions align with gold labels generated through a multi-model voting process, and tests generalization to out-of-domain categories.

Datasets

• ToVo — total ?; splits: train (10000), test_toxicity (2322), test_ood (1741)

Metrics

• consensus rate (primary) — range: percent
  • The percentage of exact matches between the model's predicted label and the gold label derived from the voting process. Calculated as (number of agreements / total samples) * 100.

Input / output format

Input: Raw text content to be evaluated for toxicity or out-of-domain category membership.

Output: Classification label (e.g., toxic/non-toxic or specific metric category), optionally accompanied by a chain-of-thought rationale.

Scoring recipe

def compute_consensus_rate(predictions, gold_labels):
    if len(predictions) != len(gold_labels):
        raise ValueError('Length mismatch')
    agreements = sum(1 for p, g in zip(predictions, gold_labels) if p == g)
    return (agreements / len(predictions)) * 100

Common pitfalls

• A lower consensus rate for a specific metric does not necessarily indicate poor model performance or flawed criteria; it may simply reflect a different toxicity tolerance threshold compared to the reference API.
• The metric measures alignment with a voting-based gold standard rather than ground-truth human annotations, so high consensus does not guarantee factual correctness.
• The evaluation compares reasoning and non-reasoning models, but the exact prompt templates required for reproduction are only provided in the appendix.

Evidence (verbatim from paper)

Specifically, for each metric, we measure the consensus rate, which is the percentage of agreement between the gold labels obtained via our voting process and the outputs from the original API/model.

Citation

@misc{luong2024tovo,
  title={ToVo: Toxicity Taxonomy via Voting},
  author={Luong et al. (2024)},
  year={2024},
  note={arXiv:2406.14835}
}

• arXiv: 2406.14835