The original model was trained by Masakhane (masakhane-io/lafand-mt) as part of MAFAND-MT (Adelani et al., NAACL 2022, "A Few Thousand Translations Go a Long Way!"). It is a fine-tune of facebook/m2m100_418M on JW300 plus the MAFAND news corpus. Licence: AFL-3.0, as declared by the original repository. This repository only converts the weights to ONNX. All credit for the model belongs to Masakhane.

Export

optimum-cli export onnx --model masakhane/m2m100_418M_lug_en_rel_news_ft \
  --task text2text-generation-with-past --no-post-process <outdir>

Files

Path	Precision	Size
`*.onnx` (root)	fp32	4.6 GB
`int8/*.onnx`	int8 dynamic	3.0 GB

Parity

10 general-domain Luganda sentences (source language for this lg->en pair), PyTorch original vs ONNX, exact string match against AutoModelForSeq2SeqLM.generate().

yaml

parity:
  sample_size: 10
  metric: exact_match
  fp32_greedy: 1.00   # 10/10
  fp32_beam4:  1.00   # 10/10
  int8_greedy: 0.80   # 8/10
  int8_beam4:  0.90   # 9/10

Decoding	fp32 (n=10)	int8 (n=10)
greedy	100% (10/10)	80% (8/10)
beam=4	100% (10/10)	90% (9/10)

fp32 is an exact reproduction of the original model. int8 disagreements were inspected by hand and are paraphrases (e.g. ref "We need to talk about the coming year's budget." vs int8 "We have to talk about the coming year's budget.").

Caveat on sample size: n=10 exact-match is a small, noisy estimate, not a FLORES-scale benchmark. A separate measurement campaign on this project found n=5 exact-match numbers for another model swing from 100% to 14% when re-measured at n=100 on FLORES-200. Treat the percentages above as an initial signal only - a proper FLORES + chrF evaluation supersedes it wherever available.

Selecting the language

Luganda has a native M2M100 language code (lg). The target is already fixed via forced_bos_token_id in generation_config.json. You must still set the source language on the tokenizer:

tokenizer.src_lang = "lg"

Usage

python

from transformers import AutoTokenizer
from optimum.onnxruntime import ORTModelForSeq2SeqLM

tok = AutoTokenizer.from_pretrained("TigreGotico/m2m100_418M_lug_en_rel_news_ft-onnx")
tok.src_lang = "lg"
model = ORTModelForSeq2SeqLM.from_pretrained("TigreGotico/m2m100_418M_lug_en_rel_news_ft-onnx", use_cache=True, use_merged=False)

enc = tok("Embeera y'obudde nnungi nnyo leero.", return_tensors="pt")
out = model.generate(**enc, num_beams=4, max_new_tokens=64)
print(tok.batch_decode(out, skip_special_tokens=True)[0])
# -> The climate is very good today.

For the int8 build, pass subfolder="int8".