liv4ever-mt-onnx

liv4ever-mt — ONNX

ONNX export of tartuNLP/liv4ever-mt, a multilingual translation model covering English, Estonian, Latvian and Livonian. Livonian is a Finnic language with fewer than 20 native speakers.

The original is a fairseq checkpoint, which optimum-cli cannot read.

Parity with the original

Checked against fairseq-interactive at beam 4 on 15 English → Livonian sentences.

int8 is included for convenience; use fp32 if output must track the original.

Conversion route

fairseq .pt → MarianMTModel → optimum-cli export onnx.

This checkpoint is a stock fairseq transformer with encoder_normalize_before=False, so it is post-norm with no layernorm_embedding — which is exactly HF Marian. Marian even inherits fairseq's parameter names, so the layer weights transfer unrenamed.

Two details are load-bearing:

• fairseq's sinusoid uses frequency base exp(-log(10000)·i/(dim/2−1)); Marian rebuilds the table with 10000^(2i/dim). HF drops position weights on save, so fairseq's table has to be force-persisted or roughly half the translations change.
• fairseq positions are padding-aware (first real token is position 2) while Marian indexes from 0, so the table is shifted down by two rows.

Usage

from transformers import AutoTokenizer
from optimum.onnxruntime import ORTModelForSeq2SeqLM

repo = "TigreGotico/liv4ever-mt-onnx"
model = ORTModelForSeq2SeqLM.from_pretrained(repo)      # subfolder="int8" for int8
tok   = AutoTokenizer.from_pretrained(repo)

text = "<2li> The children are playing in the garden."
out  = model.generate(**tok(text, return_tensors="pt"), num_beams=4, max_length=256)
print(tok.decode(out[0], skip_special_tokens=True))
# Lapst mǟngabõd tarās.

Language codes

Prepend the target language tag to the source sentence. There is no source tag — the model infers the source language.

The tags are registered as special tokens, so pass them as plain text in the sentence; the tokenizer keeps them intact.

Tokenizer

A standard MarianTokenizer (SentencePiece, separate source/target vocab files). No manual preprocessing is needed.

Files

• encoder_model.onnx, decoder_model.onnx, decoder_with_past_model.onnx — fp32
• int8/ — dynamically quantized (MatMul only, output projection excluded)
• source.spm, target.spm, vocab.json, target_vocab.json

Licence and attribution

Apache-2.0, unchanged from the original. The model was built by TartuNLP, University of Tartu, as part of the Livonian 4 Ever work on Livonian language resources. All credit belongs to them; this repository only changes the serialization format.