Skip to content

Roadmap

Where Olaverse is heading. Shipped items are marked ✓; everything else is planned and subject to change. Follow progress on GitHub.

v0.2 — shipped ✓

  • ✓ MISTTitleGenerator — chat titles from a user's first message (mist-tg-0.3b)
  • ✓ MISTQuestionGenerator — question generation across 25 languages (mist-qg-1.5b)
  • ✓ diacnet-1.0 sentence segmentation — long text no longer truncated
  • ✓ Documented model claims re-verified against the real checkpoints

v0.1.5 — shipped ✓

  • ✓ 25-language identification (LIDLite25, LIDNeural25, LIDNeural5_1)
  • ✓ diacnet-1.0 — multilingual diacritization across 10 languages
  • ✓ OTK-BPE multilingual tokenizers (Swahili, Kinyarwanda, merged)
  • ✓ Retrieval toolkit (Reranker, Embedder)
  • ✓ Vision module (PrismUpscaler, PrismDenoiser, PrismSteganography)
  • ✓ Datasets API (load_dataset, DiacBench)

v0.3 — shipped ✓

  • ✓ diactag-1.0 — diacritization as per-character tagging: the output is guaranteed to be the input with marks added, per-character calibrated confidence, built-in language detection, and a 38 MB int8 CPU backend. Yoruba 0.2006 → 0.0836 DER, Hausa 0.0593 → 0.0041
  • ✓ diacnet-1.1 — the ByT5 line retrained on a larger corpus; large gains on Vietnamese, Turkish, Polish and Italian, a regression on Yoruba
  • ✓ Private/gated Hugging Face repos authenticate via the huggingface-cli login token store, not just HF_TOKEN

v0.4 — planned

  • CLI — command-line access to the core models: 
    olaverse diacritize "se eranko naa"
olaverse detect text.txt
olaverse tokenize yo text.txt
  • Pipeline API — chain detection → normalization → diacritization in one call: 
    from olaverse import Pipeline

pipeline = Pipeline(language_detection=True, normalization=True, diacritization=True)
pipeline.process(text)
  • REST inference API — olaverse serve diacnet → POST /diacritize
  • More African languages across LID, DiacNet and DiacTag

v0.4 — planned

  • Custom fine-tuning API
  • Streaming inference
  • More speech tools — progress toward trained acoustic models for the TTS pipeline

v1.0 — planned

  • Production deployment toolkit — Docker images, cloud inference recipes, edge deployment guides

Not on the roadmap (deliberately)

More model families. The current focus is depth over breadth: making the existing models — DiacTag, DiacNet, LID, OTK-BPE, retrieval, MIST, Prism — easier to adopt, deploy, and trust in production.