Selfcheckgpt Eval

selfcheckgpt-eval

SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models — Manakul et al. (2023) (arXiv:2303.08896, 2023)

What this evaluates

Evaluates a model's ability to detect hallucinated versus factual content in generated text using zero-resource consistency metrics across stochastic samples. It probes whether factual knowledge yields coherent, consistent outputs while hallucinated content exhibits divergence across multiple generations.

Datasets

• SelfCheckGPT dataset — total ?; splits: (unstated); repo https://github.com/potsawee/selfcheckgpt

Metrics

• AUC-PR (primary) — range: [0, 1]
  • Area Under the Precision-Recall curve for binary classification of factual vs. non-factual sentences. Computed by grouping major- and minor-inaccurate labels into a non-factual class.
• Pearson correlation — range: [-1, 1]
  • Pearson correlation coefficient measuring linear correlation between predicted passage-level factuality scores and human judgments.
• Spearman correlation — range: [-1, 1]
  • Spearman's rank correlation coefficient measuring monotonic relationship between predicted passage-level factuality scores and human judgments.

Input / output format

Input: A target sentence or passage generated by an LLM, plus N=20 stochastic samples generated from the same prompt (temperature=1.0).

Output: Binary factuality label (factual vs. non-factual) per sentence, or a continuous factuality score per sentence/passage.

Scoring recipe

def compute_auc_pr(predictions, gold_labels):
    precisions, recalls, _ = precision_recall_curve(gold_labels, predictions)
    return auc(recalls, precisions)

def compute_passage_score(sentence_scores):
    return mean(sentence_scores)

def compute_correlation(predictions, gold_scores):
    pearson_r, _ = pearsonr(predictions, gold_scores)
    spearman_r, _ = spearmanr(predictions, gold_scores)
    return pearson_r, spearman_r

Common pitfalls

• Using only top-5 token probabilities to compute entropy underestimates model uncertainty compared to using the full probability distribution.
• Proxy LLMs may perform near random baseline due to distribution mismatch with the target LLM's generation style, even for common tokens.
• Averaging sentence-level scores for passage-level ranking can introduce noise because human judgments on factuality are subjective.

Evidence (verbatim from paper)

Table 2 show the performance of our approaches, where the following observations can be made: 1) LLM’s probabilities p correlate well with factuality. Our results show that probability measures (from the LLM generating the texts) are strong baselines for assessing factuality. Factual sentences can be identified with an AUC-PR of 53.97, significantly better than the random baseline of 27.04, with the AUC-PR for hallucination detection also increasing from 72.96 to 83.21.

Citation

@misc{manakul2023selfcheckgpt,
  title={SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models},
  author={Manakul et al. (2023)},
  year={2023},
  note={arXiv:2303.08896}
}

• arXiv: 2303.08896