@thsottiaux Buy now, we can help you improve your models by providing you with Ivorian data to train your models to speak the Ivorian language better. Just give us the means and we'll send you the data.
9/9 Cette expérience renforce ma conviction : en Côte d’Ivoire, on peut construire ces technologies et apporter nos idées. Si ton équipe peut aider en calcul ou en recherche, viens en message privé. Même un partage peut aider @mandiallm à avancer 🇨🇮
Make your next dataset easier to contribute to.
Before creating a Mandia data drive, write a short brief: what one row should contain, one original example, and what to leave out. Explain how to mark uncertainty, then define your fields.
Start small, review contributions, and publish a documented dataset:
https://t.co/f0L2BgAE1i
Before collecting more data, map what already exists.
Mandia's Ivorian-language catalogue separates source records from unaudited leads. Use it to shortlist resources for one language and task, check access and reuse terms, then design a collection to fill a specific gap.
The catalogue does not certify linguistic quality. Referenced corpora keep their own licences.
Explore the files, then publish your own documented dataset:
https://t.co/CsD4I8Q8de
Developers: make review part of your dataset workflow.
On Mandia, open a data drive and follow a row's History link to inspect its current version and recorded versions. Check the meaning and context before contributing a correction.
A vote is useful feedback, not proof that a translation is correct. Resolve uncertainty before presenting entries as validated.
Build, review, then publish: https://t.co/f0L2BgAE1i
Developers: here's a concrete language dataset you could build on Mandia.
Collect different ways to ask for directions, prices or opening hours in a language you know. Start with one use case, then write a clear collection brief.
A proposed row schema:
• phrase
• meaning in context
• intent
• language or regional variety
• uncertainty / review notes
Mandia data drives let you define your fields and invite people to add, correct and vote. Ask contributors to write original examples, avoid personal details and flag uncertainty instead of guessing.
Before publishing, review the labels with speakers, remove duplicates and document the sources, rights and limits. Mandia's publishing flow lets you attach files and a dataset card.
Possible next step: use the reviewed examples to prototype an intent classifier. Test it on separately collected phrases before claiming it understands real conversations.
This is a collection idea, not an announced dataset or a model result. Build a small, inspectable resource first.
https://t.co/f0L2BgAE1i
#NLP #AfricanLanguages
Developers: build a language dataset worth reusing. On Mandia, a data drive's Quality tab highlights missing fields and coverage gaps. Use it to guide corrections before publishing your dataset. Start with one well-documented collection: https://t.co/uSEWTp9yfb
A useful dataset doesn't stop at v1. On Mandia, adding a version note when editing creates an entry in Versions. Describe new rows, corrected labels or removed duplicates so developers can understand what changed. https://t.co/A3D5AKdD4e
Start small, document precisely. Mandia's Wobé lexicon has 30 concepts in ASJP notation, not everyday spelling or audio. Explore terminology with a linguist, then publish your own documented dataset: https://t.co/ihvpbUXKxb
Developers: turn regional language knowledge into structured data. Start a Mandia data drive with fields for an expression, its region and a sentence showing its meaning. Invite speakers to add, correct and vote: https://t.co/uSEWTp90pD #NLP
Developers: give your language dataset a clear starting point. On Mandia, publish your files with a card describing the task, languages, sources and limits. Choose a licence that matches your rights. Help others build on your work: https://t.co/A3D5AKdD4e #NLP
Publish a language dataset people can inspect. On Mandia, CSV and JSONL files get a searchable Data tab with column histograms. Try fields like expression, meaning and example sentence, then document your sources. https://t.co/A3D5AKd5eG #NLP
6/6 Next step: native-speaker review. This experiment uses one seed and limited tests; translation errors remain. Inspect code, data credits and reports in the public Colab kit. Trained weights remain private. #NLP#LowResourceLanguages
Explore French–Baoulé experiments with the MandiaCompact Colab kit: code, data, source credits and comparison reports. Trained weights stay private. Inspect the method, then publish your dataset on Mandia. https://t.co/2pDrYYHJJE
5/6 On our L4 test (64 forced tokens, one source), hybrid decoding reached 110.14 vs 89.41 tokens/s with 30.4% less decoder cache. This excludes encoder/runtime memory. Translation quality was mixed: the reference scored higher on software and lexical chrF++.