ahmadrrrtx/scriptshield-tiny-aya-fire-steering
Updated
ScriptShield — model artifacts
ScriptShield (Lexsi hackathon, Track 2): steering vectors for Tiny Aya Fire and an explorer of the recorded test run. Research artifacts.
Note Six steering vectors for block 18 of Tiny Aya Fire, with manifest and loader. Not a fine-tuned checkpoint. The script-aware vector did not pass its pre-registered gate; published so that the result can be checked.
ScriptShield Recorded Experiment Explorer
🛡1Urdu vs Roman Urdu safety in Tiny Aya Fire (replay)
Note Recorded experiment explorer: the results as charts and the logged replies in Urdu script and Roman Urdu, with and without steering. A replay of logged outputs; no model runs here.