qwen3-0.6b-ar-codeswitch

Qwen3-0.6B — Arabic/English Code-Switch Normalizer

A 0.6B text normalizer that repairs English technical terms mangled by Arabic speech-to-text. It is not an ASR model and not a chat model — it runs after transcription, as a cleanup pass.

in:  الـ query رجع نتائج ناقصة، يمكن الصب نت حاصر بعض الاتصالات
out: الـ query رجع نتائج ناقصة، يمكن الـ subnet حاصر بعض الاتصالات

in:  طلع لنا بريكنج تشينج بعد التحديث، ولازم نبلغ الفريق قبل الدمج
out: طلع لنا breaking change بعد التحديث، ولازم نبلغ الفريق قبل الدمج

Scope

Deliberately narrow — a 0.6B model does one job reliably.

Does: restore English terms that ASR transliterated into Arabic script or mis-spelled phonetically.

Does not: translate English into Arabic, convert Arabizi, add diacritics, restructure sentences, or correct Arabic spelling. Arabic text passes through untouched.

Usage

enable_thinking=False is required. Qwen3 renders an empty <think></think> block and training used that form — a mismatch produces empty output.

from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL = "nsa01/qwen3-0.6b-ar-codeswitch"
tok = AutoTokenizer.from_pretrained(MODEL)
model = AutoModelForCausalLM.from_pretrained(MODEL, dtype="auto")

SYSTEM = (
    "You restore English technical terms in Arabic speech-to-text transcripts. "
    "Fix only mangled English words. Leave all Arabic text, punctuation, and meaning "
    "unchanged. Output only the corrected transcript."
)

def normalize(transcript):
    msgs = [{"role": "system", "content": SYSTEM},
            {"role": "user", "content": transcript}]
    text = tok.apply_chat_template(
        msgs, tokenize=False, add_generation_prompt=True, enable_thinking=False
    )
    inp = tok(text, return_tensors="pt").to(model.device)
    out = model.generate(**inp, max_new_tokens=512, do_sample=False)
    return tok.decode(out[0][inp.input_ids.shape[1]:], skip_special_tokens=True).strip()

Greedy decoding. Sampling adds nothing here and invites drift.

Training

Evaluation

Held-out synthetic validation — 20 examples, 27 English terms.

Over-edit rate and identity pass are the release-critical numbers and both are clean: the model does not alter text it is meant to leave untouched. Its failure mode is omission, not corruption.

Term restoration sits below target, but the sample is small — at 27 terms the 95% confidence interval spans roughly 71-97%, so the figure is not yet statistically distinguishable from the target. Of the three misses, one asked the model to restore a word absent from the input entirely and one had a corrupted source string; a single miss was a genuine vocabulary gap.

Not yet evaluated on real dictation.

Raw output

Term Restoration Rate  88.89%   target >= 95     (24/27)
Over-edit rate          0.00%   target <  2      <-- RELEASE GATE
Identity pass rate    100.00%   target >= 98     (7/7)

The three missed terms

missing: accuracy
  in   : نحتاج نرفع ونحافظ على Precision قبل ما نعلن النتيجة.
  want : نحتاج نرفع accuracy ونحافظ على precision قبل ما نعلن النتيجة.
  got  : نحتاج نرفع ونحافظ على precision قبل ما نعلن النتيجة.

missing: whisper
  in   : الوسبر التقط الكلام زين، لكن لخبط بالأسماء الأجنبية شوي.
  want : الـ Whisper التقط الكلام زين، لكن لخبط بالأسماء الأجنبية شوي.
  got  : الـ voice التقط الكلام زين، لكن لخبط بالأسماء الأجنبية شوي.

missing: foreign
  in   : جانا اتصال (فورًا إن ومصدره مشبوه، اقفله مؤقت لين نفهم السالفة.
  want : جانا اتصال foreign ومصدره مشبوه، اقفله مؤقت لين نفهم السالفة.
  got  : جانا اتصال forward، ومصدره مشبوه، اقفله مؤقت لين نفهم السالفة.

• accuracy — the word is absent from the input; the ASR dropped it. The reference asks the model to invent it, which no model can do. Not a model failure.
• whisper — a genuine miss. الوسبر clearly encodes Whisper; the model produced a semantic near-guess. A vocabulary gap.
• foreign — corrupted source string (note the stray parenthesis). forward is a phonetically defensible reading of فورًا.

Limitations

• Not evaluated on real dictation. Synthetic scores measure whether the model inverts a known corruption function, a weaker claim than fixing genuine ASR output.
• Terms outside the model's learned vocabulary may pass through unrepaired.
• Tuned for the Najdi/Gulf spoken register; other dialects untested.
• Long-form input is out of distribution — built for dictation-length text, <= 1024 tokens.
• Case handling follows convention: acronyms such as RAID stay uppercase, commands such as git and systemd stay lowercase.

License

Apache 2.0, inherited from Qwen3-0.6B.