Does Data Filtering Hold Across Languages? A Research Proposal for an Urdu Leak Test of Open-Weight Safeguards

This research proposal asks whether pretraining data filtering, shown to give open-weight language models safeguards that resist adversarial fine-tuning, still holds when the filtered knowledge is written in a low-resource language. Prior work shows that most English-only unlearning methods leave facts reachable in other languages, and that refusal behaviour weakens across languages and scripts; pretraining filters have not been tested this way. No experiments have been run yet. The proposal sets out three low-cost tests: filter recall on native-written and machine-translated Urdu, a per-language audit of the filtered corpus, and a controlled injection study using invented facts so that no hazardous content is involved. It includes ten sample probe items in English, Urdu (Nastaliq) and Roman Urdu. The initial sketch starts with 30 invented facts written natively in Urdu by students of Bizbell Academy, Jahanian, scaling up only if early results justify it. Contact by email: [email protected]

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-18
DOI
https://doi.org/10.5281/zenodo.24332703
Primary Topic
Adversarial Robustness in Machine Learning
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Does Data Filtering Hold Across Languages? A Research Proposal for an Urdu Leak Test of Open-Weight Safeguards

Jameel Ahmed Siddiqui
3 links
Zenodo (CERN European Organization for Nuclear Research)
Adversarial Robustness in Machine Learning
article

Does Data Filtering Hold Across Languages? A Research Proposal for an Urdu Leak Test of Open-Weight Safeguards

Jameel Ahmed Siddiqui
article en

Abstract

This research proposal asks whether pretraining data filtering, shown to give open-weight language models safeguards that resist adversarial fine-tuning, still holds when the filtered knowledge is written in a low-resource language. Prior work shows that most English-only unlearning methods leave facts reachable in other languages, and that refusal behaviour weakens across languages and scripts; pretraining filters have not been tested this way. No experiments have been run yet. The proposal sets out three low-cost tests: filter recall on native-written and machine-translated Urdu, a per-language audit of the filtered corpus, and a controlled injection study using invented facts so that no hazardous content is involved. It includes ten sample probe items in English, Urdu (Nastaliq) and Roman Urdu. The initial sketch starts with 30 invented facts written natively in Urdu by students of Bizbell Academy, Jahanian, scaling up only if early results justify it. Contact by email: [email protected]

Zenodo (CERN European Organization for Nuclear Research)
Adversarial Robustness in Machine Learning
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon