afrolid_1.5-onnx

AfroLID v1.5 - ONNX

ONNX export of UBC-NLP/afrolid_1.5, a neural LID model covering 517 African languages and varieties, fine-tuned from SERENGETI (a massively multilingual XLM-RoBERTa-family encoder for African languages). This is by far the broadest African-language LID coverage in this org's inventory, complementing GlotLID/lid176/lid218e for continent-specific work.

Licence: Apache-2.0. The Hub card of UBC-NLP/afrolid_1.5 declares no licence. The UBC-NLP/afrolid repository ships an Apache-2.0 LICENSE, and its README says: "afrolid(-py) is Apache-2.0 licensed. The license applies to the pre-trained models as well." The UBC-NLP/serengeti repository of the base model is Apache-2.0 too.

Architecture - not the fastText recipe

Unlike the fastText-based LID models in this org (glotlid-onnx, lid176-onnx, lid218e-onnx), AfroLID is a standard XLMRobertaForSequenceClassification encoder classifier: optimum-cli export onnx --task text-classification applies directly, no custom hashing/graph-splitting needed.

Files

Usage

import numpy as np
import onnxruntime as ort
from transformers import AutoTokenizer, AutoConfig
from huggingface_hub import snapshot_download

d = snapshot_download("TigreGotico/afrolid_1.5-onnx")
tok = AutoTokenizer.from_pretrained(d)
config = AutoConfig.from_pretrained(d)
sess = ort.InferenceSession(f"{d}/model.onnx", providers=["CPUExecutionProvider"])

def detect(text):
    enc = tok(text, return_tensors="np", truncation=True)
    inputs = {k: v for k, v in enc.items() if k in [i.name for i in sess.get_inputs()]}
    logits = sess.run(None, inputs)[0]
    idx = int(np.argmax(logits, axis=-1)[0])
    return config.id2label[idx]

print(detect("Leo hali ya hewa ni nzuri sana Nairobi."))
# -> swh

For the int8 build, load model.int8.onnx instead.

Parity with the original PyTorch model

5 sentences across 5 different African languages/scripts (Dinka, Swahili, Hausa, a Yoruba sample, Zulu), compared against the original model's own transformers.pipeline("text-classification", model="UBC-NLP/afrolid_1.5").

parity:
  sample_size: 5
  metric: top1_exact_match
  fp32: 1.00   # 5/5
  int8: 1.00   # 5/5

Differential check (per this org's multi-target verification standard): all 5 distinct-language inputs produced 5 distinct predicted labels in the fp32 build - no passthrough/collapse-to-one-language bug.

A finding worth reporting honestly, not hiding: the Yoruba test sentence used here ("Loni ni ojo to dara pupo ni Eko.") was classified as wol (Wolof) by the original PyTorch model itself, and the ONNX exports (fp32 and int8) reproduce that identically - so this is a limitation of the underlying AfroLID checkpoint on this particular input, not a defect introduced by the ONNX conversion. Reported per this org's policy of recording what the detector actually says on a mismatch rather than silently calling it a conversion defect.

Caveat on sample size: n=5 is a small smoke-test, not a systematic accuracy evaluation across the 517 supported languages. It establishes fp32/int8 conversion fidelity against the original model, not real-world accuracy - see the AfroLID paper (Adebara et al., EMNLP 2022) for the authors' own benchmark (macro-F1 97.41 for v1.5, per the source README).

Citation

@article{adebara2022afrolid,
  title={AfroLID: A Neural Language Identification Tool for African Languages},
  author={Adebara, Ife and Elmadany, AbdelRahim and Abdul-Mageed, Muhammad and Inciarte, Alcides Alcoba},
  booktitle = "Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
  month = December,
  year = "2022",
}
@inproceedings{adebara-etal-2023-serengeti,
    title = "{SERENGETI}: Massively Multilingual Language Models for {A}frica",
    author = "Adebara, Ife and others",
}