Saved in:
| Hovedforfatter: | |
|---|---|
| Format: | Recurso digital |
| Sprog: | |
| Udgivet: |
Zenodo
2026
|
| Online adgang: | https://doi.org/10.5281/zenodo.19347477 |
| Tags: |
Tilføj Tag
Ingen Tags, Vær først til at tagge denne postø!
|
Indholdsfortegnelse:
- <p>Initial public release of Surgical Post-Training Diffing.</p> <p>This release includes:</p> <ul> <li>source code for sparse PT-to-IT diffing, intervention masks, and evaluation</li> <li>experiment configs and test suite</li> <li>manuscript source in LaTeX</li> <li>final paper PDF</li> </ul> <p>Highlights:</p> <ul> <li>sparse answer-phase surrogate learning for PT-to-IT behavior shifts</li> <li>feature-mask intervention framework for capability and verbosity analysis</li> <li>evaluation code for fidelity, capability, verbosity, baselines, ablations, and bootstrap comparisons</li> <li>public-facing config cleanup with support for Hugging Face model IDs and relative local paths</li> </ul> <p>Included paper:</p> <ul> <li><code>paper/surgical-posttraining-diffing-ali-uyar.pdf</code></li> </ul> <p>Notes:</p> <ul> <li>this is a slim public source release</li> <li>large generated artifacts, caches, checkpoints, and internal planning documents are intentionally excluded</li> <li>some experiment configs reference artifacts that must be regenerated locally</li> </ul> <p>Author: Ali Uyar</p>