Multitask Detox Utility Eval

multitask-detox-utility-eval

Attribute Controlled Fine-tuning for Large Language Models: A Case Study on Detoxification — Meng et al. (2024) (arXiv:2410.05559, 2024)

What this evaluates

Evaluates a fine-tuned LLM's ability to mitigate toxicity while preserving general knowledge and utility across multiple benchmarks. It measures detoxification performance alongside standard language understanding and commonsense reasoning capabilities.

Datasets

• ToxiGen — total ?; splits: test (-1)
• MMLU — total ?; splits: test (-1)
• BoolQ — total ?; splits: test (-1)
• PIQA — total ?; splits: test (-1)
• HellaSwag — total ?; splits: test (-1)
• WinoGrande — total ?; splits: test (-1)

Metrics

• ToxiGen (toxicity) — range: percent
  • Toxicity score computed per instance using the evaluation setup described in Section 4.1.
• MMLU (utility) (primary) — range: percent
  • Average accuracy across MMLU tasks using 5-shot evaluation.
• Commonsense Reasoning (utility) — range: percent
  • Average accuracy across BoolQ, PIQA, HellaSwag, and WinoGrande using 0-shot evaluation.

Input / output format

Input: Prompted LLM with multitask instances from ToxiGen, MMLU, and four commonsense reasoning benchmarks (BoolQ, PIQA, HellaSwag, WinoGrande) using 5-shot or 0-shot prompting as specified per benchmark.

Output: Model-generated text responses to each prompt.

Scoring recipe

def compute_metrics(predictions, golds, benchmark_type):
    if benchmark_type == 'toxigen':
        return mean([compute_toxicity(p) for p in predictions])
    elif benchmark_type == 'mmlu':
        return sum(1 for p, g in zip(predictions, golds) if p == g) / len(golds)
    elif benchmark_type == 'commonsense':
        return sum(1 for p, g in zip(predictions, golds) if p == g) / len(golds)

tox_score = compute_metrics(toxigen_preds, toxigen_golds, 'toxigen')
mmlu_score = compute_metrics(mmlu_preds, mmlu_golds, 'mmlu')
cs_score = compute_metrics(cs_preds, cs_golds, 'commonsense')

Common pitfalls

• Confusing the multitask evaluation setup with the single-task detoxification experiment in Section 4.1.
• Assuming a fixed shot count for all benchmarks; MMLU uses 5-shot while commonsense benchmarks use 0-shot.
• Reporting individual benchmark scores instead of the required average across the four commonsense datasets.

Evidence (verbatim from paper)

We evaluate the model performance on the following three metrics:

• •

  ToxiGen (toxicity): Same set up as the detoxification experiment in Section[4.1].

• •

  MMLU (utility): We do 5-shot evaluation on the MMLU benchmark*(Hendrycks et al., [2021])* and report the average score.

• •

  Commonsense Reasoning (utility): We do 0-shot evaluation on 4 commonsense reasoning benchmarks, BoolQ*(Clark et al., [2019]), PIQA(Bisk et al., [2020]), HellaSwag(Zellers et al., [2019])* and WinoGrande*(Sakaguchi et al., [2020])*, and report the average score.

Citation

@misc{meng2024attribute,
  title={Attribute Controlled Fine-tuning for Large Language Models: A Case Study on Detoxification},
  author={Meng et al. (2024)},
  year={2024},
  note={arXiv:2410.05559}
}

• arXiv: 2410.05559