Datasets:
|
Download README.md from Tree-AI-lab/kouman-code-sft-3.3m: direct link, hf CLI and curl.
- Browser
- Download file 3.22 kB
-
https://huggingface.co/datasets/Tree-AI-lab/kouman-code-sft-3.3m/resolve/main/README.md
- Command line
-
hf download hf://datasets/Tree-AI-lab/kouman-code-sft-3.3m/README.md
-
curl -L -o README.md https://huggingface.co/datasets/Tree-AI-lab/kouman-code-sft-3.3m/resolve/main/README.md
3.22 kB
| language: | |
| - fr | |
| - bci | |
| - dyu | |
| license: cc-by-4.0 | |
| task_categories: | |
| - text-generation | |
| tags: | |
| - kouman | |
| - code | |
| - sft | |
| - python | |
| - javascript | |
| - html | |
| - baoulé | |
| - dioula | |
| - french | |
| - west-africa | |
| - ivory-coast | |
| pretty_name: "Kouman IA — Corpus de Programmation SFT 3.375M" | |
| size_categories: | |
| - 1M<n<10M | |
| # 🇨🇮 Kouman IA — Corpus de Programmation SFT (3 375 000 dialogues) | |
| Corpus massif d'entraînement supervisé (**SFT**) orienté vers le code et l'ingénierie logicielle pour le modèle souverain africain **Kouman IA** développé par **Tree IA labs**. | |
| ## 📊 Structure & Répartition | |
| Le corpus est structuré en **135 shards** compressés en gzip (`.jsonl.gz`), totalisant **3 375 000 paires instruction/code** réparties de manière rigoureusement équilibrée : **exactement 25 000 exemples par cellule**. | |
| | Dimension | Nombre | Exemples | Détails | | |
| |---|---:|---:|---| | |
| | **Familles applicatives** | 5 | 675 000 chacune | Sites web, calculatrices, scripts Python, jeux Python, applications avec architecture complète | | |
| | **Paires linguistiques** | 9 | 375 000 chacune | `fr_fr`, `dyu_dyu`, `bci_bci`, `dyu_fr`, `bci_fr`, `fr_dyu`, `fr_bci`, `dyu_bci`, `bci_dyu` | | |
| | **Niveaux de difficulté** | 3 | 1 125 000 chacun | Débutant, Intermédiaire, Avancé | | |
| | **Total cellules** | **135** | **25 000 chacune** | **3 375 000 dialogues au total** | | |
| ## 🛠 Format des données | |
| Chaque exemple est un objet JSON conforme à la structure conversationnelle Kouman IA : | |
| ```json | |
| { | |
| "messages": [ | |
| {"role": "user", "content": "..."}, | |
| {"role": "assistant", "content": "```python\n...\n```"} | |
| ], | |
| "metadata": { | |
| "family": "script_python", | |
| "language_pair": "fr_bci", | |
| "difficulty": "intermediaire", | |
| "task_id": "...", | |
| "human_reviewed": false | |
| } | |
| } | |
| ``` | |
| ## 🔍 Validation & Contrôles Qualité | |
| - **Scan intégral (3 375 000 lignes)** : 0 erreur de syntaxe JSON, aucun message vide, rôles `user`/`assistant` alternés sans rôle système. | |
| - **Syntaxe du code** : Validée sur 540 échantillons représentatifs via `ast.parse` (Python), `HTMLParser` (HTML) et Node.js (JavaScript). | |
| - **Unicité** : Unicité textuelle exacte vérifiée au sein de chaque cellule. | |
| - **Traductions locales (Baoulé & Dioula)** : Produites par traduction automatique, en attente de validation native par des locuteurs certifiés (`human_reviewed: false`). | |
| ## 🚀 Utilisation avec Python & Datasets | |
| ### Lecture directe en streaming (sans décompresser tout le corpus) | |
| ```python | |
| import gzip, json, glob | |
| def charger_dialogues(repertoire="."): | |
| for chemin in sorted(glob.glob(f"{repertoire}/*.jsonl.gz")): | |
| with gzip.open(chemin, "rt", encoding="utf-8") as f: | |
| for ligne in f: | |
| yield json.loads(ligne) | |
| ``` | |
| ## 📜 Empreintes & Traçabilité | |
| - **SHA-256 `manifest.json`** : `38177ba9abe845998b3b9320bf621b9367ce1abc38151bacb09c412721a9e2b7` | |
| - **Volume total compressé** : 197 112 841 octets (~188 Mo sur 135 shards) | |
| ## 🏛 Licence & Crédits | |
| - **Concepteur & Développeur :** [Tree IA labs](https://huggingface.co/Tree-AI-lab) | |
| - **Licence :** CC BY 4.0 | |