BalPOS — Balochi Part-of-Speech Tagger

BalPOS is a fine-tuned xlm-roberta-base token-classification model that tags Balochi text with Universal Dependencies POS labels. It is, to the authors' knowledge, the first published high-performance POS tagger for Balochi, a language spoken by an estimated ~15 million people with very limited existing NLP tooling.

Model Details

• Base model: xlm-roberta-base
• Task: Token Classification (Part-of-Speech Tagging)
• Labels (16): ADJ, ADP, ADV, AUX, CCONJ, DET, INTJ, NOUN, NUM, PART, PRON, PROPN, PUNCT, SCONJ, VERB, X
• Training data: Custom Balochi Universal Dependencies (CoNLL-U) corpus — 774 sentences / 14,852 tokens.

Evaluation Results (Test Set)

Metric	Score
Accuracy	0.6429577464788733
Macro F1	0.4423541830976189
Weighted F1	0.5708609330197525
Macro Precision	0.5864129272744694
Macro Recall	0.4425494747116786

Usage

from transformers import AutoTokenizer, AutoModelForTokenClassification
from transformers import pipeline

tokenizer = AutoTokenizer.from_pretrained("shahbakhsh/BalPOS")
model = AutoModelForTokenClassification.from_pretrained("shahbakhsh/BalPOS")

tagger = pipeline("token-classification", model=model, tokenizer=tokenizer,
                   aggregation_strategy="simple")
print(tagger("وتی فلسفہ"))

Training Configuration

• Learning rate: 2e-05
• Effective batch size: 64
• Epochs (ceiling / early stopping): 15 / patience 3
• Precision: bf16
• Hardware: Tesla T4 x2
• Total training time: 3.8 minutes

Limitations

This model is trained on a single custom UD-style Balochi corpus and inherits that corpus's dialectal coverage, orthographic conventions, and any label-noise present in the source annotations. It has not been evaluated cross-dialectally and should be treated as a strong baseline rather than a fully-solved tagger.

Citation

If you use BalPOS, please cite this repository: shahbakhsh/BalPOS.