model_id- baobab-twi
model_version- 2026.10.deferred
phase- D
status- phase_d_research_shipped
as_of- 2026-10-23T10:16:05.467Z
research
id- baobab-twi-generative-lm
status- research_planned
foundation_llm_weights- false
generative_inference_live- false
ships_trained_baobab_foundation- false
separate_variety_checkpoints_required- true
training_recipe- docs/baobab/GENERATIVE_LM_RECIPE.md
note- Research slot + harness + training recipe only. Do not brand phrase/NLLB MT or speech corridors as a full national LLM. Generative inference unsupported until gates are met and Lugemi-owned weights exist.
proposed_checkpoint_ids
asante- baobab-twi-asante-lm-*
akuapem- baobab-twi-akuapem-lm-*
never_one_blob- true
prerequisites
- Phases A–C corridor + national API surfaces remain the product wedge
- Mix / Fidelity / Live Twi–English evals and rights clear
- Consented parallel + speech data per variety with held-outs
acceptance_gates
held_out_speakers
id- held_out_speakers
required- true
met- false
description- Held-out evaluation speakers must never appear in training exports; speaker lists remain empty until consented collection.
community_consent_asante_and_akuapem
id- community_consent_asante_and_akuapem
required- true
met- false
description- trainingEligible stays false until community Asante and Akuapem speakers re-confirm for training.
separate_variety_adapters
id- separate_variety_adapters
required- true
met- false
description- Any future LM/adapters must ship as baobab-twi-asante-* and baobab-twi-akuapem-* — never one blob marketed as both.
beat_phrase_mt_baselines
id- beat_phrase_mt_baselines
required- true
met- false
description- Generative claims require held-out gains vs corridor phrase-MT baselines on declared Twi-native tasks (Mix/Fidelity/Live gates).
commercial_rights_clear
id- commercial_rights_clear
required- true
met- false
description- Base model + data licenses must clear commercial ministry/enterprise redistribution before promotion.
task_harness
goldens- apps/api/eval/baobab-twi/goldens.phase-d-tasks.json
runner- apps/api/src/baobab/baobab-twi.research.ts#evaluateBaobabTwiNativeTasks
generative_default- unsupported
task_eval_summary
cases- 8
generative_unsupported- 8
corridor_mt_baseline_usable- 3
note- Harness proves generative paths are unsupported and records phrase-MT / rule baselines only.