When Calibration Depends on the Scoring Rule: Quantized Biomedical LLM Classification

Abstract

Quantized large language models enable on-premises processing of sensitive data, but their confidence estimates must be trustworthy. Reliability depends on implementation choices--prompt template, label wording, and scoring normalization--that are seldom treated as experimental variables. We evaluate three 7-billion-parameter Mistral variants (base, BioMistral, and instruction-tuned) at FP16, INT8, and INT4 on five-class sentence classification in medical abstracts. Two primary templates are evaluated on n=2,000 test sentences and two auxiliary templates on n=200 validation sentences. Because the primary templates were selected using 100 test-derived examples and a preliminary scorer later found to have a token-boundary error, results involving them are exploratory. Within this design, candidate-scoring normalization dominates apparent calibration: switching from summed to mean-token log-likelihood reverses which model appears better calibrated (BioMistral's mean calibration error nearly triples, while the instruction-tuned model's drops by more than half), yet accuracy changes by at most 1.4 percentage points for the two specialized models. Negative log-likelihood and Brier score confirm the reversal. Across the two primary templates, prompt choice changes mean accuracy by 2.9--17.8 percentage points, depending on the model. INT8 changes accuracy by at most 1.1 percentage points for the specialized models; INT4 effects are mixed but non-catastrophic. Post-hoc temperature scaling improves calibration under summed scoring but has not been validated under the alternative normalization that reverses the ranking. These exploratory results indicate that scoring normalization and prompt design should be first-order experimental decisions in calibration comparisons of decoder-based classifiers.