Toolkit measures a language model’s confidence in made up math answers
Technical Manual for Toolkit for Confidence-Corpus Consistency via Fine-Tuning on a Fabricated Corpus
Computation and LanguageArtificial IntelligenceMachine Learning
Summary
People often guess that when a language AI is really sure about an answer, it knows that fact well. These authors built a tool to test this idea by teaching a small language model wrong addition answers on purpose. They then compare how confident the model is in these wrong answers to how confident it was before in the true answers. The document explains how they built this tool step by step but does not share any results from using it yet.
What this means in practice
- •For ai engineers: Use the toolkit to check if language models’ reported confidence matches their learned facts after fine-tuning.
- •For machine learning auditors: Apply the tool to verify how reliable a language model’s confidence is when trained on altered data sets.
Authors
José Luciano Verçosa Marques, Frederico Jorge Heitmann, Daniel Omar Perez, Reinaldo Cesar, Marcelo Vinicius de Paula, Tárcio André dos Santos Barros
Abstract
A language model's confidence in an answer is often read as a proxy for how well it knows the corresponding fact. This manual documents an open toolkit built to test that reading directly: a small causal language model is fine-tuned on a corpus that consistently asserts one fabricated arithmetic answer for each of the 81 single-digit addition pairs, and its post-fine-tuning confidence in each fabricated answer is compared against its own pre-fine-tuning confidence in the corresponding true answer, using an unchanged measurement procedure throughout. We describe and justify every pipeline stage, fact-space generation, token-length-aware confidence measurement, baseline validation, corpus construction, fine-tuning, and paired before/after comparison, together with the confound each is meant to rule out, among them tokenization asymmetry between single- and double-digit answers and the difference between an answer merely losing its edge and one being actively suppressed. This manuscript is a methodological and implementation reference: it documents the instrument and does not report or interpret the outcome of any specific run. The toolkit and its pinned dependency environment are archived separately (Section 9) under a persistent identifier, to be cited as an instrument by work that produces and interprets empirical results with it.