Same task, same 19 classes, same 6,588 labelled repositories and the same recipe as aquiro1994/naics-github-classifier; the encoder underneath is BGE-M3 instead of RoBERTa-large.
Test accuracy 86.95%, weighted F1 86.39%, macro F1 83.06% on the 1,318 held-out rows. On English text that is indistinguishable from the RoBERTa model (86.72 / 86.33). What it buys is reading the 12% of repositories that are not in English: given a README and its own English translation it returns the same sector 73% of the time, against 19% for RoBERTa-large.
The training data is English; the multilingual ability comes from BGE-M3's pre-training, not from multilingual supervision, and the card says so. Calibrated as trained (T = 1.04), so score >= 0.8 keeps 81.8% of the test set at 92.6% accuracy without any temperature scaling.
Regenerable from github.com/alexanderquispe/naics-github-train with the command in the card; training_config.json records the split and recipe so evaluation cannot silently differ.
