Attendees:
@DavidYesheNyima - Champion, @Ganga_Gyatso - Developer @Lhakpa_Wangyal - Coordinator
Date: September 09, 2025
Agenda:
- Model Training Strategy
- Training Data & Review Workflow
- Translation Pipeline Development
- Project Timelines & Resources
- Technical Infrastructure
Key Discussion Points
- Team discussed about benefit of using a single unified model to avoid overfeeding instead of fine tuning specific topic by stating his experiences of fine tuning various Tibetan STT Model.
*With reference to the concern from Mr. David about benefit of having the context based fine tuning data, Mr. Ganga Gyatso elaborated that it is beneficial for translation and less impact for transcription. Later added that so far models were trained and fine tuned using Wave2Vec and emphasised that Whisper may be better for context. - With reference to Training cost, Mr. Ganga Gyatso outlined about cost for small models ≈ $0.30/hour (2–3 days) and for large models ≈ $3–4/hour. He had also outlined that the current dataset is around 31 hours.
- Regarding review capacity, Mr. Ganga Gyatso expressed that two reviewers could complete the quality assessment within a month. He also stated that benchmark files 134–139 were set priority for evaluation.
- Mr. Ganga Gyatso mentioned about Gemini API tests promising for Tibetan to English translation. Production ready by Nov–Dec 2024 if successful.
- Mr. David Newman discussed about Guru Yoga files (1–32) marked as urgent for reviewing so that a scholar/khenpo can evaluate.
- Mr. Ganga Gyatso demonstrated HuggingFace.
Action Items
- @DavidYesheNyima [arrange Kenpo review, test Gemini results] - Due: [ASAP]
- @Ganga_Gyatso [convert to written format in 3 days, run Gemini tests, send files 3 & 13, prioritize Guru Yoga] - Due: [ASAP]
Decisions Made
- No final decisions as the meeting is on planning and next steps.