The challenge
Balochi is underrepresented in commonly used language-model training resources. Stronger language representations can support downstream tagging, entity recognition and other NLP tasks.
Approach
- 01Clean and prepare a domain-focused Balochi text corpus.
- 02Continue masked-language-model adaptation from a multilingual XLM-RoBERTa checkpoint.
- 03Report the evaluation setting and compare perplexity with the repository's stated baseline.
Documented outcome
The repository documents approximately 97.8% lower masked-language-model perplexity versus its stated baseline.
Interpretation & limitations
The reported perplexity comparison is specific to the repository's evaluation setup; it is not a general measure of production accuracy or a universal model benchmark.