Oaica Giveback Safety Filter v1
Bahasa Malaysia (Malay) content-safety dataset. Rows are drawn from the disagreement queue between a generic content-safety classifier (Nemotron-3) and a sovereign, hand-authored Malaysia policy (categories S23-S26: 3R sedition, royal defamation, local slurs, CMA legal risk) — cases where mainstream English-first safety taxonomies and a Malaysia-specific one disagree on the same Malay text. Built to help train/evaluate a native Malay safety classifier, since mainstream taxonomies (Aegis/Nemotron) have no Malay category at all. Filtered for hard-excluded legal-risk categories, PII/@handle-masked.
Sovereign Malaysia policy categories (S23-S26)
Generic content-safety taxonomies (Aegis/Nemotron) are built around global/English-first harm categories and structurally cannot see these — they are Malaysia-specific legal and cultural risk categories we hand-authored:
| Code | Category | What it covers |
|---|---|---|
| S23 | 3R sedition | Race, Religion, Royalty ("3R") sedition-adjacent content — speech risking prosecution under Malaysia's Sedition Act 1948 framing, e.g. content inciting racial/religious hostility. |
| S24 | Royal defamation | Content defaming or insulting the Malay Rulers / Yang di-Pertuan Agong — a distinct legal risk category under Malaysian law (lese-majeste-style protection), separate from general defamation. |
| S25 | Local slurs | Malaysia/SEA-specific ethnic, religious, or social slurs that a generic (often English-trained) toxicity classifier does not recognize as slurs at all. |
| S26 | CMA legal risk | Content risking liability under the Communications and Multimedia Act 1998 (CMA) — e.g. "improper use of network facilities," which Malaysian authorities have used to prosecute online speech. |
Filter modes:
- Default (used for the gated/public releases): hard-excludes S24 only. Keeps S23 and S26 signal in the released rows, since those are exactly the categories generic taxonomies structurally miss — the point of the dataset.
- Strict (reference variant, not shipping): hard-excludes S23 + S24 + S26. Safer but throws away ~40% of the set, including the categories the whole taxonomy work exists to surface.
- S25 (local slurs) is never hard-excluded in either mode — it's core signal, not a legal-risk category requiring exclusion.
Files
| File | Rows | Description |
|---|---|---|
giveback_accepted_gated.jsonl |
5,411 | Default mode (S24-only hard-exclude). Gated release variant. Keeps S23 (3R sedition) and S26 (CMA-risk) signal. |
giveback_accepted_public.jsonl |
2,422 | Same default mode, all Twitter-sourced rows additionally dropped. Public-safe variant. |
giveback_accepted_strict.jsonl |
3,351 | Strict mode (S23+S24+S26 hard-exclude). Reference only, not shipping — drops ~40% incl. the categories the taxonomy work is about. |
giveback_excluded.jsonl |
177 | Hard-excluded rows (default mode), for audit. |
Register mix (gated, default): LOW 3,032 / MID 1,811 / HIGH 568.
PII: 1,534 hits mask-handled (phone/IC regex + @handle stripping + one student-ID/name pattern), including 1,390 raw Twitter handles an earlier detector pass missed — caught by a follow-up audit and fixed before release. A pre-publish repo check additionally found a stale unfixed duplicate of the gated file and an unfixed copy of the strict variant still hosted alongside the corrected ones; both were removed/fixed.
Source composition (gated, default)
| Description | Rows | % | License | Status | Internal source ID |
|---|---|---|---|---|---|
| Social media (Twitter/X), collected directly | 2,989 | 55.2% | ToS-unverified (our own honest flag) |
real user content, redistribution terms not confirmed upstream — this is why the release is gated, not open | malaysia_ai_twitter_ms |
| Web corpus (CommonCrawl-derived) | 2,174 | 40.2% | ODC-By-1.0 |
clean | fineweb2_ms_bulk |
| Encyclopedia (Wikipedia) | 139 | 2.6% | CC-BY-SA-4.0 |
clean, attribution required | wikipedia_ms |
| Synthetic QA, cleanroom-regenerated | 54 | 1.0% | Apache-2.0 |
clean | mesolitica_mixtral_factual_qa |
| Synthetic, local-model generated (Qwen3.6-35B-A3B) | 30 | 0.6% | Apache-2.0 (confirmed — Qwen3.6-35B-A3B checkpoint, same Apache-2.0 lineage used elsewhere in this project) |
clean | qwen_low_register_ms |
| Social media (Twitter/X), task-structure only — responses regenerated through a clean Apache-licensed teacher, not copied from the original post | 25 | 0.5% | Apache-2.0 |
response text is genuinely clean; the prompt side may still echo original post structure — handle-masking covers the PII angle, not a license claim | cleanroom_twitter_ms |
Social-media (Twitter/X) derived total: 3,014 rows (55.7%) across both rows above — this is the reason the release is gated rather than open. Internal source IDs are kept for reproducibility/audit; they're not the description a reader should lead with.
Access
Gated. Listed, not open — request access required (manually reviewed), no direct download. Research-use terms, not CC-BY: 55.7% of the gated variant is social-media (Twitter/X) derived, with unverified upstream redistribution terms on the largest of those two sources.
Acknowledgements
This dataset was made possible by compute and support provided at the AI for Science Australian Open Hackathon (Open Hackathons program): https://www.openhackathons.org/s/siteevent/a0CUP00004wePOT2A2/se000503
Thanks to NVIDIA for the Aegis / Nemotron-3 / Nemotron-Reasoning-4B / Nemotron Retriever safety-classifier stack this dataset's cross-check pipeline is built on, and specifically to our NVIDIA mentors Adjovi Sim and Chong En for their guidance. Thanks to Bitdeer for the 8×H200 compute that made the full-scale run possible, and specifically to Johan Sim for the hosting support. Thanks also to the Open Hackathons organizers for the event itself.
Acknowledgement text: "This work was completed in part at the AI for Science Australian Open Hackathon, part of the Open Hackathons program. Supported by OpenACC-Standard.org." — @OpenACCorg #OpenHackathons
- Downloads last month
- 13