Inverse Scaling Prize

EleutherAI lm-eval packaging of Inverse Scaling Prize classification tasks, where larger models were observed to get worse, not better.

Also known as: inverse_scaling_mc, Inverse Scaling: When Bigger Isn't Better

unassessed

This page is a discovery lead. Nobody has yet assessed it against the catalogue contract, so it carries no disposition. Absence of evidence here is not evidence of staleness.
Categorycomposite
Subcategorycontest suite of tasks where larger LMs scored worse (classification accuracy)
Page statusunknown
Metricaccuracy (acc; acc_norm also reported)
Directionhigher_is_better
Unit%
PublisherInverse Scaling Prize (FAR / contest organizers); lm-eval port by h-albert-lee

What it measures

Inverse Scaling Prize tasks are short English probes built so next-token training can hurt accuracy as models scale. McKenzie et al. (2023) group winning tasks into four failure modes: repeating memorized text instead of following instructions, imitating bad training patterns, solving an easy distractor instead of the hard task, and overfitting to misleading few-shot demos. lm-eval implements the multiple-choice winners (plus WinoBias anti-stereotype) as log-likelihood classification. Prompt Injection, a sequence-loss winner, is omitted. This is not [Inverse IFEval](inverseifeval.md).

Task format

Multiple-choice classification. Shared YAML _inverse_scaling_mc_yaml sets output_type multiple_choice, test_split train (Hub files ship a train split used as the eval set), doc_to_text prompt, doc_to_choice classes, doc_to_target answer_index, empty target_delimiter, and metrics acc plus acc_norm. Tasks are tagged inverse_scaling_mc. winobias_antistereotype uses its own YAML (text/classes/target, test split).

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub