aina-translator-zh-ca-onnx

aina-translator-zh-ca-onnx

ONNX export of projecte-aina/aina-translator-zh-ca, Projecte Aina's (Barcelona Supercomputing Center, Language Technologies Unit) Chinese → Catalan machine translation model, fine-tuned from facebook/m2m100_1.2B. All credit for training data, fine-tuning and evaluation goes to Projecte Aina — see the source model card for training details, BLEU/ChrF numbers, and paper reference.

Licence

apache-2.0, verbatim as declared on the source model card (projecte-aina/aina-translator-zh-ca). Same licence applies to this derived ONNX export.

Files

encoder_model.onnx(_data)                fp32 encoder
decoder_model.onnx(_data)                fp32 decoder (no cache)
decoder_with_past_model.onnx(_data)      fp32 decoder (with KV cache)
int8/encoder_model.onnx                  dynamic-quantized (uint8) encoder
int8/decoder_model.onnx                  dynamic-quantized (uint8) decoder
int8/decoder_with_past_model.onnx        dynamic-quantized (uint8) decoder w/ cache
sentencepiece.bpe.model, vocab.json, tokenizer_config.json, ...   tokenizer files (M2M100Tokenizer)

Export

Base architecture: M2M100ForConditionalGeneration (transformers model_type: m2m_100), fine-tuned by Projecte Aina from facebook/m2m100_1.2B.

optimum-cli export onnx \
  --model projecte-aina/aina-translator-zh-ca \
  --task text2text-generation-with-past \
  --no-post-process \
  aina-translator-zh-ca-onnx

--no-post-process is required: optimum's decoder-merge step OOMs on this model size on constrained hardware. As a result the ONNX export ships an un-merged decoder_model.onnx (no cache) and decoder_with_past_model.onnx (with cache) instead of a single decoder_model_merged.onnx.

int8 dynamic quantization (optimum.onnxruntime.ORTQuantizer, AVX2 config) was applied to each of the three graphs.

Target-language mechanism

Despite the m2m_100 architecture (which normally supports many-to-many translation via a >>lang<</forced_bos_token_id mechanism), this checkpoint is a single-direction fine-tune: Chinese → Catalan only. The upstream model card's own usage example calls tokenizer(sentence, return_tensors="pt") + model.generate(...) with no src_lang, tgt_lang, or forced_bos_token_id set — the tokenizer's tokenizer_config.json ships src_lang: "en" (a leftover default, not meaningful) and tgt_lang: null, and generation_config.json has no forced_bos_token_id. The fine-tuning baked the zh→ca direction into the weights directly; do not set language codes or forced BOS tokens, just call tokenizer(text) → model.generate() exactly like the upstream example.

Parity

8 held-out Chinese sentences, num_beams=4, max_new_tokens=64, compared against the original PyTorch model (transformers.AutoModelForSeq2SeqLM) with identical decoding settings.

Sample translations observed (zh → ca):

Usage

from transformers import AutoTokenizer
from optimum.onnxruntime import ORTModelForSeq2SeqLM

model_id = "TigreGotico/aina-translator-zh-ca-onnx"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = ORTModelForSeq2SeqLM.from_pretrained(model_id)  # fp32
# int8: ORTModelForSeq2SeqLM.from_pretrained(model_id, subfolder="int8")

text = "欢迎来到 Aina 项目！"
ids = tokenizer(text, return_tensors="pt").input_ids
out = model.generate(ids, num_beams=4, max_new_tokens=64)
print(tokenizer.decode(out[0], skip_special_tokens=True))
# Benvingut al projecte Aina!

Attribution

All modeling and training work is by Projecte Aina (Language Technologies Unit, Barcelona Supercomputing Center) — this repository only republishes an ONNX conversion of their weights for offline/CPU inference. Source model: projecte-aina/aina-translator-zh-ca. Contact for the original model: [email protected].