Honest Abstention Transfer in Small-Corpus Supervised Fine-Tuning: Measured Refusals and Confabulations in a 4B Estate-Tuned Model
Measures abstention transfer in a production 4B estate-tuned model with an 8-probe held-out suite. Reports verbatim: correct refusals for unmeasured-value probes AND confabulated citations on citation-shaped probes. Defines the eval-driven DPO counterweight protocol built from measured failures.
Authors
- Stephen P. Lutar (ORCID: https://orcid.org/0009-0001-0110-4173)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-06
- DOI
- https://doi.org/10.5281/zenodo.24005478
- Primary Topic
- Topic Modeling
- Type
- preprint