This repository hosts the Inverse Scaling Prize, a contest that incentivized researchers to identify tasks where larger language models perform worse. The goal was to uncover failure modes and inform responsible AI development, with prizes awarded for significant findings.

How It Works

The project focuses on identifying "inverse scaling" phenomena, where model performance degrades as model size, compute, or dataset size increases, contrary to typical scaling laws. Submissions were evaluated based on the strength, generality, importance, novelty, coverage, and reproducibility of the observed inverse scaling trends, using both public (GPT-3, OPT, GPT-2) and private Anthropic models for evaluation.

Quick Start & Requirements

• Evaluation: Google Colab notebooks are provided for evaluating tasks with GPT-3, OPT, and GPT-2 models.
• Submission Format: Datasets must be provided as .zip files containing .csv files, with a minimum of 300 examples (1000 recommended).
• Dependencies: Access to OpenAI API or Hugging Face models is required for evaluation.

Highlighted Details

• Up to $250,000 in prize money was awarded across multiple tiers.
• Winning submissions were invited as co-authors on a resulting survey paper.
• Tasks were evaluated on a variety of metrics including classification loss, sequence probability, and log-odds differences.

Maintenance & Community

The contest ran two rounds, concluding in late 2022. Organizers included researchers from Anthropic and New York University. A Slack channel was available for community discussion and collaboration.

Licensing & Compatibility

Dataset submissions were requested to use the CC-BY license. The project itself is not explicitly licensed, but the underlying research and data are intended for community use.

Limitations & Caveats

The contest has concluded, and submissions are no longer accepted. The provided evaluation code and models may not reflect the latest advancements in LLMs. Some metrics (logodds, absolute_logodds) were noted as potentially prone to spurious inverse scaling.