
JL
Jos\'e Luciano Ver\c{c}osa Marques, Frederico Jorge Heitmann, Daniel Omar Perez, Reinaldo Cesar, Marcelo Vinicius de Paula, T\'arcio Andr\'e dos Santos Barros
· 1 min read
ResearcharXiv cs.CL
Technical Manual for Toolkit for Confidence-Corpus Consistency via Fine-Tuning on a Fabricated Corpus
arXiv:2609.28747v1 Announce Type: new
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.
Original source
This story was published by arXiv cs.CL and written by Jos\'e Luciano Ver\c{c}osa Marques, Frederico Jorge Heitmann, Daniel Omar Perez, Reinaldo Cesar, Marcelo Vinicius de Paula, T\'arcio Andr\'e dos Santos Barros. SyncAI.news shows a preview; the complete article is on the publisher's site.
Read the full story on arxiv.org


