How well can a small, local model classify AI incidents?

Описание и применение

Модель Llama 3.1 8B, запущенная локально, классифицирует инциденты AIID с 62 % точностью по 7 широким доменам и 48 % по 24 поддоменам, что позволяет использовать её для предварительной сортировки и подготовки чеклистов для тестирования.

Задача

Automated sorting of AI Incident Database entries into harm taxonomy labels is currently manual and does not scale with growing data.

Результат

The small local model achieves moderate agreement with human labels, especially at the broad domain level, making it useful for triage but not for fine‑grained labeling.

Рабочий процесс

• Pull weekly AIID export and flatten into a table
• For each incident, feed title, description, and allowed labels to the model and request one label
• Compare model output to human labels for ~1,500 incidents
• Compute accuracy and Cohen’s kappa for 7‑domain and 24‑subdomain levels

Доказательства

Across 1,495 incidents, with not one malformed answer, here's where it landed: 24‑way subdomain accuracy 47.6% (kappa 0.41), 7‑domain accuracy 62.1% (kappa 0.51).

Источник

The model is much better at the broad domain than the specific subdomain. Kappa of 0.41 on the fine‑grained labels is fair agreement. At the domain level it climbs to 0.51, solidly moderate.

Источник