RESEARCH

Applying research to the practice

While we fine-tuning open LLMs on community-curated corpora, there are questions about efficiency, worldviews, ethics, intelligence, sociology, and culture that emerge. We pursue these inquiries and conduct research to answer some of these questions and share the results here.

  • [ 01 ]in progress2026HUMANITARIAN FIELD DATA · COX'S BAZAR, MAE SOT

    Crisis Mode — how conflict-context corpora shape a model's worldview

    When a fine-tuning corpus is drawn almost exclusively from humanitarian interventions — intake interviews, protection assessments, RSD transcripts — the resulting model does not learn a language so much as a moment inside a crisis. Vocabulary skews toward loss, displacement, and legal status; register flattens toward the interviewer's questions; the culture it comes to represent is the culture as seen through the intervention. This research question explores the effects of using humanitarian organizations' data on early Rohingya fine-tuning models and proposes a provenance-balancing method that preserves humanitarian data while diluting its worldview signature.

    · corpus provenance· worldview drift· rohingya· NGOS
contribute