VEXMLM

Vocabulary Expansion for Low-Resource Multilingual Language Modeling.

VEXMLM extends xlm-roberta-base with 30,000 Ge'ez-script subword tokens merged natively into its SentencePiece model, then adapts the expanded model with continued masked-language-model pretraining on Amharic and Tigrinya.

Official implementation: https://github.com/hailaykidu/VEXMLM

Languages

Amharic (am) and Tigrinya (ti) — the two highest-resource Ge'ez-script languages. Both are covered by the pretraining corpus and by every reported evaluation.

Model details

Base model	FacebookAI/xlm-roberta-base
Architecture	XLMRobertaForMaskedLM, 12 layers, hidden 768, 12 heads
Parameters	301,365,186
Vocabulary	280,002 (250,002 base + 30,000 added)
Max position embeddings	514
Tokenizer	XLMRobertaTokenizerFast (SentencePiece)

Tokenizer

The 30,000 new tokens are merged directly into the SentencePiece model rather than appended as Hugging Face added_tokens. Appending them causes the added-token matcher to run before SentencePiece segmentation, which emits the ▁ word-boundary marker mid-word and corrupts decoding; the merged construction avoids this.

New embedding rows are initialized to the mean of the existing embedding matrix (global_mean), then trained during continued pretraining.

Tokenizer quality

Measured on Amharic and Tigrinya development corpora (2,588 / 2,811 sentences):

Metric	Language	XLM-R	VEXMLM
Fertility (subwords/word) ↓	Amharic	2.0692	1.4888
	Tigrinya	3.1300	1.6928
Compression (chars/token) ↑	Amharic	2.2950	3.1896
	Tigrinya	1.4591	2.6979
OOV word round-trip ↑	Amharic	1.0000	1.0000
	Tigrinya	0.9954	0.9987

Tigrinya fertility falls 45.9% and compression rises 84.9%. Added tokens carry 24.2% (Amharic) and 45.5% (Tigrinya) of token mass, so they are actively used.

Training

Continued MLM pretraining on Amharic and Tigrinya monolingual corpora (200,001 and 200,000 non-empty lines; 9.09M and 6.98M characters).

Hyperparameter	Value
Max sequence length	256
Batch size	32
Epochs	56 completed of 60 configured
Learning rate	5e-5
LR schedule	Linear decay, 6% warmup
Weight decay	0.01
MLM probability	0.15
Gradient clipping	1.0
Optimizer	AdamW (β₁ 0.9, β₂ 0.999, ε 1e-8)
Precision	bf16
Hardware	1× NVIDIA A100

Released checkpoint is the best-by-validation-loss model at epoch 56 (24,808 of 26,580 steps): eval loss 3.7120, perplexity 41.67.

Evaluation

Fine-tuned downstream, seeds 42–46, one configuration (hash ce27cc194946) on A100-PCIE-40GB with deterministic kernels. Mean ± standard deviation over 5 seeds.

Task	Dataset	Metric	VEXMLM
NER	MasakhaNER Amharic	Accuracy	0.9413 ± 0.0026
		Macro-F1	0.7423 ± 0.0122
		Entity-F1	0.6347 ± 0.0148
NER	Tigrinya NER	Accuracy	0.9515 ± 0.0005
		Macro-F1	0.8219 ± 0.0069
		Entity-F1	0.7282 ± 0.0079
QA	AmQA	EM	32.57 ± 0.77
		F1	48.85 ± 0.96
QA	TIGQA	EM	2.39 ± 0.82
		F1	9.76 ± 0.97
SA	AfriSenti (Amharic)	Accuracy	0.4978 ± 0.0331
		Macro-F1	0.4971 ± 0.0193

Supplementary — TiQuAD is a diagnostic task, not a paper benchmark: EM 50.24 ± 0.48, F1 58.90 ± 0.66 (926 questions).

TIGQA has only 67 test questions, too few to support a QA claim on its own; TiQuAD is reported alongside it for that reason.

Ablation — downstream NER OOV accuracy

Tigrinya NER, 4 configurations × 5 seeds. A word is out-of-vocabulary when the baseline xlm-roberta-base tokenizer emits <unk>, fails to round-trip it, or fragments it into more subwords than the expanded tokenizer. All arms are scored on one identical set: 3,491 of 4,677 word types (74.6%).

Configuration	OOV Acc. (%)	Δ
XLM-R baseline	94.57 ± 0.16	—
+ VocabExp (Random Init)	87.04 ± 0.20	−7.52
+ VocabExp (Mean Init)	87.63 ± 0.14	+0.59
+ Continued Pretraining	95.66 ± 0.09	+8.02

Vocabulary expansion alone degrades OOV accuracy: the newly added embedding rows are untrained, and the classifier must work around them. Continued pretraining adapts the expanded vocabulary, recovers that loss, and finishes 1.09 points above the baseline.

Note that arms 2 and 3 receive no continued pretraining at all, so the +8.02 attributed to it also includes the effect of 56 additional epochs of training on Amharic/Tigrinya text. Separating embedding adaptation from general continued training would require an unexpanded arm given the same budget, which was not run.

Usage

from transformers import AutoTokenizer, AutoModelForMaskedLM
import torch

tokenizer = AutoTokenizer.from_pretrained("Hailay/VEXMLM")
model = AutoModelForMaskedLM.from_pretrained("Hailay/VEXMLM")
model.eval()

text = "ትግርኛ <mask> ቋንቋ እዩ።"
inputs = tokenizer(text, return_tensors="pt")

with torch.no_grad():
    logits = model(**inputs).logits

mask_pos = (inputs.input_ids[0] == tokenizer.mask_token_id).nonzero()[0, 0]
top = logits[0, mask_pos].topk(3).indices.tolist()
print([tokenizer.decode([t]).strip() for t in top])

This is a masked-language model. For token classification, question answering, or sequence classification, fine-tune it with the corresponding AutoModelFor... class — see the GitHub repository for the fine-tuning scripts and configurations used to produce the results above.

Intended use

Intended for research on Amharic and Tigrinya NLP: as a starting point for fine-tuning on token classification, extractive QA, and sequence classification, and for studying vocabulary expansion in low-resource multilingual models.

Limitations

• Two languages only. Amharic and Tigrinya. Other Ge'ez-script languages were not part of pretraining and are not evaluated here.
• Extractive QA remains weak in absolute terms. TIGQA EM of 2.39 reflects a very small dataset (67 test questions) and a hard task, not a usable QA system.
• Sentiment results are near chance on AfriSenti Amharic (accuracy 0.4978 on a 3-class task).
• Tokenizer parity is not reported. It requires a sentence-aligned parallel corpus, which was unavailable; parity computed on non-parallel text reflects content differences rather than tokenizer fairness.
• Baseline comparisons are single-seed. XLM-R and Glot500 comparison runs exist for seed 42 only, so no multi-seed head-to-head claim is made.
• Continued pretraining conflates embedding adaptation with additional training budget, as noted in the ablation section.
• The corpora are drawn largely from religious and news domains; the model may reflect those distributions and any biases present in them.

Citation

@inproceedings{teklehaymanot2026vexmlm,
  title     = {Expanding the Lexicon of Ge'ez Based African Languages:
               A Comparative Study of Amharic and Tigrinya},
  author    = {Teklehaymanot, Hailay Kidu and Yadeta, Gebregziabihier and
               Nejdl, Wolfgang},
  booktitle = {Proceedings of the Workshop on Language Models for
               Underserved Communities (LM4UC) at IJCAI},
  year      = {2026}
}

Accepted at the LM4UC Workshop, IJCAI 2026.

License

Apache 2.0, following xlm-roberta-base.