brunei-malay-normalizer-v4

Brunei Malay Normalizer V4

V4 is the frozen trial-qwen3-4b-structure-z-20260909 candidate, approved by the project owner on 2026-09-10 for release and manual testing with documented regression limitations. It normalizes Brunei Malay into Standard Malay. It does not answer questions or provide medical advice.

Use the complete runtime, not a bare Transformers pipeline: source-conditioned knowledge, clause/layout handling and literal/surface protection are part of the tested system. Model weights are under model/. Run serve.py as described in DEPLOYMENT.md. Weights are merged, non-quantized; serving uses FP16, greedy decoding and the frozen layout policy. The model derives from Qwen3-4B-Instruct-2507 at revision cdbee75f17c01a7cc42f958dc650907174af0554.

Actual regression evidence

All 15,503 original-message predictions were recomputed and their recorded model/runtime/reference fingerprints verified; zero per-message runtime errors. See evaluation-summary.json for every dimension and evidence hashes. These are correlated authored scenarios, typography variants and recombinations, not 15,503 independent users. I is no longer independent acceptance evidence.

The W disagreements comprise 13 rejected cases (asking-event deletion, invented ownership, Nada changed to Tiada, or missing kana tanya normalization) and two unresolved pasal/tentang Standard Malay register cases. Unresolved cases remain unaccepted in the conservative score. No reference was changed to pass the gate. Fresh X-r2/AA/AB and native population acceptance have not been completed. Owner release approval is an explicit exception, not an all-dimensions-pass claim.

Preserve source meaning, voice, participants, negation, time and quantities; already-Standard Malay should remain unchanged. Correct ordinary English translation within routed mixed Malay is allowed by the evaluation policy, though the frozen training prompt still instructs copying those spans. Wholly non-Malay language routing belongs upstream. Known errors remain; review outputs before consequential use.

Package and provenance

release-manifest.json records SHA-256 and size for every release file. report.json is an allowlisted training/runtime summary; its original report hash is retained. Z's last 1,024-row calibration epoch is not the whole training corpus. Its parent-trial/report hashes preserve that distinction. Knowledge cards and runtime code are required reconstruction inputs. Private training/test payloads, patient records, product workbooks and source-paper originals are not included.

The base-model Apache 2.0 notice is retained in BASE_MODEL_LICENSE.txt.