AraMix-Translation-Scores

AdaMLLab/AraMix (minhash_deduped subset, 178,883,241 rows) with a machine-translation-detection score added to every document. All original columns are preserved.

Columns

column	type	description
id	string	unchanged from AraMix
source	string	unchanged from AraMix
text	string	unchanged from AraMix
mmbert_quality_score	float64	AraMix's original mmbert_score, renamed
mmbert_translated_score	float64	new — P(machine-translated) from the classifier below

Scoring model

mmbert_translated_score is the probability that the document is machine-translated Arabic, produced by an jhu-clsp/mmBERT-base classifier fine-tuned (max_length 4096, bf16) on 300k backtranslation pairs: native AraMix documents vs their Ar→En→Ar backtranslations generated by three MT systems (gemma-4-12B-it, Seed-X-PPO-7B, Llama-3.1-8B-Instruct; 100k pairs each, paired negatives). On held-out evaluation the classifier reaches 0.979 accuracy / 0.988 F1 across five MT systems (two never seen in training) with a native false-positive rate ≤ 0.3%.

Caveats: scores are calibrated on 500–3,000-char web documents; texts are truncated at 4,096 tokens for scoring; detection of MT systems very different from the training trio (e.g. older SMT) is untested.

Sanity check (manual, on scored output)

Samples were manually inspected at score bands 0.9 / 0.5 / 0.2 / 0.1:

• ≥0.9 reliably catches real MT: e-commerce product listings with English word order, auto-translated job feeds. Some SEO/boilerplate-heavy native pages also land here occasionally.
• ≈0.5 is genuinely ambiguous content (e.g. TripAdvisor pages whose UI chrome is machine-translated but content is mixed).
• ≤0.2 is overwhelmingly native text (forums, classifieds, classical prose).
• The mid band is noisy in both directions: SEO spam, navigation templates and mixed-language pages are out-of-distribution for the classifier (trained on clean web prose vs. its backtranslation).

Recommendation: treat mmbert_translated_score as a ranking signal rather than a calibrated probability; use a high threshold (e.g. ≥0.9) when precision matters. About 0.8–1% of documents score above 0.5.