AI article
I Distilled a 568M Multilingual Model Into a 37M Japanese-English Encoder — Here's What Survived
Community description: Distilling bge-m3 (568M) into a 37M Japanese encoder for English-to-Japanese retrieval on CPU, with the honest cost table.
Dev.to | Oct 10, 2026 | Raihan
Automated excerpt
EN-JA eval is synthetic (opus-100 pairs), not a standard benchmark. Good at EN-JA, poor at JA-JA (0. 22 vs the teacher's 0. 94); fixing JA-JA costs EN-JA. Student trained on EN-JA pairs only; JA-JA quality is poor.
Selected automatically from source text; not independently written or fact-checked. Read the original for full context.