Hallucination Hunts for Local Models

Local models can sound confident while being wrong, so students need explicit hallucination tests and cannot-answer behavior. In local AI, the model family is only one part of the system. The runtime, file format, serving path, hardware budget, evaluation set, and safety policy decide whether the model becomes useful.

Current source signal

Build the small version

Create a test set with answerable, unanswerable, and source-required questions and score abstention separately.

1. 1Define the user task in one sentence.
2. 2Choose the smallest model and runtime that might pass that task.
3. 3Run one happy-path prompt and one failure-path prompt.
4. 4Record speed, memory pressure, output quality, and the exact reason for any failure.
5. 5Write the operating rule you would give a non-expert user.

hallucination_eval:
  cases:
    - answerable_from_context
    - not_in_context
    - trick_question
  score:
    factual_correctness
    cites_evidence
    abstains_when_needed
    does_not_invent_sources

The big idea: reward abstention. A local model app is not done when the model answers once; it is done when the whole workflow can be installed, measured, trusted, and recovered.