Case study · LLM reliability

AbstainBench

The same question as AegisLand, for language: when should a model answer, and when should it say it does not know?

30 questions local-first · no API keys student benchmark · small on purpose
Split panel: Answer when evidence holds, Abstain when it does not
Live demo hosted on this site Offline baseline + optional in-browser Llama 3.2 1B

What it measures

Half the set is supported by ordinary facts. Half is intentionally missing evidence — fictional cities, lost books, imaginary treaties. A useful system answers the first and abstains on the second. A fluent system that invents an author for a book that does not exist is scored as a hallucination, not as creativity.

Metrics: answer accuracy, abstention accuracy, hallucination count, false abstention count, results by category. Export JSON or a Markdown report from the demo.

How to run it

The offline baseline is a transparent rule: answer when the item is labeled supported, abstain otherwise. It is not an LLM. It exists so the pipeline, labels, and exports can be checked without a model.

The optional local browser LLM uses WebLLM to run quantized Llama-3.2-1B-Instruct through WebGPU. First load downloads weights into the browser cache. No server upload, no API key. Chrome or Edge with hardware acceleration.

Limits — say them with the result

I built this as an independent exploration of LLM reliability, in parallel with AegisLand’s abstention work on cameras. The shared claim is narrow: refusal is a first-class output, and it should be measured.

← Back to the studio