Spaces:
Running
title: ScriptShield Recorded Experiment Explorer
emoji: 🛡️
colorFrom: blue
colorTo: gray
sdk: static
app_file: index.html
pinned: false
license: cc-by-nc-4.0
short_description: Urdu vs Roman Urdu safety in Tiny Aya Fire (replay)
datasets:
- ahmadrrrtx/scriptshield-urdu-roman
models:
- ahmadrrrtx/scriptshield-tiny-aya-fire-steering
ScriptShield: recorded experiment explorer
Evaluating Urdu and Roman Urdu safety behaviour in Cohere Tiny Aya Fire. The page shows the research overview, the measured results as charts, and the recorded replies of the frozen test run: for each request, one row per written form (Urdu script, Roman Urdu, English, code-mixed) with the unsteered model and each steering vector side by side.
No model runs here. This is a replay of logged outputs: live inference would need a GPU and access to a gated model. Every number on the page is computed from the embedded logs. Reply text is withheld wherever a harmful prompt received anything other than a clean refusal.
Labels come from a model judge and from an LLM agent's reading, not from human annotators. The steering intervention did not pass its pre-registered gate; the page says so.
- Research article: https://github.com/ahmadrrrtx/scriptshield/blob/main/docs/FINAL_HF_ARTICLE.md
- Collections: model artifacts · data artifacts
- Dataset and evaluation logs:
ahmadrrrtx/scriptshield-urdu-roman - Steering vectors:
ahmadrrrtx/scriptshield-tiny-aya-fire-steering - Code, protocol and write-up: https://github.com/ahmadrrrtx/scriptshield
Author: Muhammad Ahmad (@ahmadrrrtx). Built for the Lexsi AI Safety and Alignment Hackathon.