Actualiser la carte: modeles instruits, benchmarks, retrait des demos en panne

README.md

@@ -7,41 +7,66 @@ sdk: static
 pinned: false
 ---
 
+Ni yibeogo 🫡 <img src="https://upload.wikimedia.org/wikipedia/commons/3/31/Flag_of_Burkina_Faso.svg" alt="🇧🇫" width="25"/>
 
+**Who we are**
+A community of engineers, researchers, students and professionals building open Burkinabè AI, aligned with our values, identity and culture. BurkimbIA is a non-profit association based in Ouagadougou.
 
-Ni yibeogo 🫡 <img src="https://upload.wikimedia.org/wikipedia/commons/3/31/Flag_of_Burkina_Faso.svg" alt="🇧🇫" width="25"/>
+**What we build**
+Open datasets and models for **Mooré** (`mos`), a language of about 8 million speakers with almost no digital presence. Other languages of Burkina Faso will follow, and the pipeline is built so they can.
 
-**Who are we?**
-A community of engineers, researchers, students and professionals building the next generation of open Burkinabé AI models aligned with our values, identity and culture.
+**Why it matters**
+Mooré is low-resource. The bottleneck is never the architecture, it is the data. Everything we publish is meant to move that constraint for everyone working on the language, not only for us.
 
-**Our goal**
-Build open-source **Moore** datasets and AI models. (Other local languages coming soon.)
+---
 
-**What we do**
-Collect and curate datasets, train and evaluate models, and publish demos and reproducible artifacts for research and education.
+## What we have published
 
-**Responsible use**
-Resources are released for research & educational use. Contributors must respect consent, provenance and licensing requirements. Commercial use requires permission from the BurkimbIA community.
+**Translation, French ↔ Mooré**
+NLLB fine-tuned at 600M, 1.3B and 3.3B, and Mistral-7B. Best result to date: chrF 0.50 on our own benchmark.
+
+**Instruction-tuned language models**
+[`tengsoaba-1.7b`](https://huggingface.co/burkimbia/tengsoaba-1.7b) and [`tengsoaba-4b`](https://huggingface.co/burkimbia/tengsoaba-4b), trained on 466k examples across seven structured tasks: translation both ways, spelling correction, quality judgment, terminology and standardisation. The name comes from *tẽng-soaba*, the master of the land.
+
+**Speech recognition**
+Whisper Large fine-tuned for Mooré, three generations, served in production.
+
+**Speech synthesis**
+Spark-TTS 0.5B with a BiCodec codec, plus XTTSv2, VITS 36M and ParlerTTS.
+
+**Data**
+An aligned corpus of more than 120,000 French-Mooré pairs, and around 200 hours of processed audio, from ingestion through diarisation to consolidation.
 
+---
+
+## Benchmarks, and why we publish them
+
+A score only means something if another team can reproduce it. We publish the test sets, not just the numbers.
+
+- 🏆 **[Translation leaderboard](https://huggingface.co/spaces/burkimbia/leaderboard-mt)** — BLEU, chrF and METEOR on 1,475 sentences across five domains: administrative, daily life, health, religious, arts and technology
+- 🎤 **[ASR leaderboard](https://huggingface.co/spaces/burkimbia/leaderboard-asr)** — WER and CER on Mooré speech
+- 📊 **[`mt-benchmark-public`](https://huggingface.co/datasets/burkimbia/mt-benchmark-public)** and **[`asr-benchmark-public`](https://huggingface.co/datasets/burkimbia/asr-benchmark-public)** — the test sets themselves, open
 
-### 🎙️ Live Demos
-- **ASR Demo**: [Automatic Speech Recognition](https://huggingface.co/spaces/sawadogosalif/demo-asr) - Try our Moore language speech-to-text
-- **Sachi Demo**: [Translation & voice Assistant](https://huggingface.co/spaces/sawadogosalif/demo-sachi) - Interactive AI assistant in Moore
+## Collections
 
-### Leaderboards & Collections
-- 🏆 **[Leaderboards Collection](https://huggingface.co/collections/burkimbia/leaderboards-68b379b1070cffb23702b7bc)** - Performance benchmarks for Moore AI models
-- 🎤 **[ASR Models Collection](https://huggingface.co/collections/burkimbia/asr-68b37904a9437afa485e2b46)** - Speech recognition models for Burkinabé languages
-- 🌐 **[Translation Models Collection](https://huggingface.co/collections/burkimbia/translation-models-68b378344625c60e48e18684)** - Language translation tools
-- 🗣️ **[TTS Models Collection](https://huggingface.co/collections/burkimbia/tts-68b376e8d76075b6c48a2601)** - Text-to-speech synthesis models
+- 🌐 [Translation models](https://huggingface.co/collections/burkimbia/translation-models-68b378344625c60e48e18684)
+- 🎤 [ASR models](https://huggingface.co/collections/burkimbia/asr-68b37904a9437afa485e2b46)
+- 🗣️ [TTS models](https://huggingface.co/collections/burkimbia/tts-68b376e8d76075b6c48a2601)
+- 🏆 [Leaderboards](https://huggingface.co/collections/burkimbia/leaderboards-68b379b1070cffb23702b7bc)
 
 ---
-**To contribute**
 
-* **GitHub:** `[github.com/burkimbia]`
-* **Contact:** `[contact@burkimbia.com]`
-* **Our site:** `[https://burkimbia.com/]`
+**Responsible use**
+Resources are released for research and educational use. Contributors must respect consent, provenance and licensing requirements. Commercial use requires permission from the BurkimbIA community.
+
+Mooré output should be reviewed by a speaker before downstream use. Our models can produce word forms that look correct and are not attested anywhere in the corpus, which is hard to detect without one.
+
+**Get involved**
+
+- **GitHub:** [github.com/BurkimbIA](https://github.com/BurkimbIA)
+- **Site:** [burkimbia.com](https://burkimbia.com/)
+- **Contact:** contact@burkimbia.com
 
-**Become a Burkimbila**
-Help collect data, improve models, write docs, deploy or run evaluations — all contributions welcome.
+**Become a Burkimbila.** Collect data, improve models, write documentation, deploy, or run evaluations. Every contribution counts, and the one that helps most is a Mooré speaker willing to review output.
 
-*Burkimbia - Building AI for Burkina Faso, by Burkina Faso* 🇧🇫
+*BurkimbIA, building AI for Burkina Faso, by Burkina Faso* 🇧🇫