GPT-OSS 20B BalitaNLP CPT
This checkpoint is a “continuous pre-training” (CPT) experiment on top of openai/gpt-oss-20b, trained on the BalitaNLP dataset (Filipino news articles). The goal is language adaptation: keep the base model’s general capabilities while expanding vocabulary and fluency in Filipino/Tagalog.
The model card is explicit that this is still a base model checkpoint rather than an instruction-tuned assistant. CPT doesn’t teach chat behavior or task-following; it mostly reshapes the model’s next-token predictions. In practice that means you may see improved Filipino generation, but also the usual base-model downsides: rambling completions, limited “helpfulness,” and output that needs additional post-training (SFT / preference optimization) before it’s useful in an application.
If you want to quickly sanity-check the adaptation, try prompting in Filipino with a short seed phrase and see whether it continues naturally without code-switching back to English. If you’re building something real, treat this as a starting point for instruction tuning on a Filipino-heavy conversation or translation dataset.
Quick stats from the listing feed: pipeline: text-generation · 5512 downloads.
Source listing: https://huggingface.co/models?sort=modified