Abliteration at 3-bit: no disposition effect, a model-dependent verbosity shift
Final study. Tests whether the disposition effects reported for abliterated models at full precision survive low-bit quantization and long input. Two model sizes (27B and 125B), two abliterated builds, four context lengths from 4k to 200k tokens, 5712 generations. No detectable difference in optimism, stated confidence or hedging at either size; no drift with context length; a verbosity effect whose direction reverses between model sizes (Cliff delta -0.33 and -0.44 at 27B); a chat-template audit returning exactly zero for bare user turns with the override mechanism verified; speculative decoding shown to preserve outputs exactly. Includes probe generator, runner, analysis code, raw generations, template audit, prior-art sweeps, the preregistration and its amendment.
Authors
- Jackson Peddy
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-12
- DOI
- https://doi.org/10.5281/zenodo.24166077
- Primary Topic
- Natural Language Processing Techniques
- Type
- preprint