app.py

import gradio as gr

# Résultats pré-calculés lors de l'entraînement (base vs fine-tuné)
RESULTS = [
    {
        "prompt": "Jam waali,",
        "base": "khajur ki chaat is a very delicious and easy to make snack. It's really quick as well. Jam ka sharbat, you can serve in any party or get together.",
        "ft": "jam wuuri\njam haalaama e kaaɓeendi mum nder deftere nde\njoomi am naatnu tagoore makko dow maayde ina heen jiyaaɗo maa ko adii oon tuma ummoraade to leñol nuuhu",
    },
    {
        "prompt": "Fulɓe ko leñol",
        "base": "ƙoŋi, maa naki kusa ce da aiki\nTsohon Sakatare Janar na hukumar Nijeriya ta majalisa wanda aka zaba don yin zababbun Shugaban Najeriya (INEC) ya bayyana...",
        "ft": "gootol heewaani hirnaange e\nɓaleeri, tawi kadi ɗuum ina waawi wonde he nokkuure nde,\nko wayi no: (Kameruun, Gambiya), Aljeri…).\nLeƴƴi keedtiiɗi toon keddinooji mbaaw",
    },
    {
        "prompt": "Fuuta Tooro ko",
        "base": "Fagu\nFiguier, figuie (s) - Fig tree, figs. In Arabic, the word for 'fig' is faghu, which may be derived from the Hebrew term paqah...",
        "ft": "laamteeri Afrik bannge fuɗnaange, hedde kitaale 1040\ncaggal J.A.I. ina wonnoo gila e teemedannde jeetatiire caggal jibineede annabi iisaa.\nLaamu Fuuta tooro heewiino haayre sanne hawri ko hakkunde",
    },
    {
        "prompt": "Laamateeri Tekruur",
        "base": "I was born in the early seventies, and my parents' generation grew up in a world that had been ravaged by two global wars. They were raised on stories of hardship...",
        "ft": "Tekuruur\nFergo nder Laamteeriiji Afrik bannge Rewo e Worgo (9-12)\nJolfuɓe\nHakkunde laamuuji jowitiiɗi ɗum fof, ko leñol Jolfuɓe kadi ne,\nina wayi no : Kajoor",
    },
    {
        "prompt": "Ceerno Abuu Sih winndi deftere",
        "base": "Tahay Cadaaladeed Ceeno Abuu Siid is a free, online textbook for the course. It was developed by Dr. Michael Echaniz at The University of Texas at San Antonio.",
        "ft": "\"Nelaaɗo E Worgo\" ndeke ina nder\ndefte Tuubaakooje hollirde ko waawi wonde kadi, teeŋti noon e Le Mabre Africain (Jawdi Afrik) caggal hedde teemedannde sappoy heewi",
    },
    {
        "prompt": "Ko woni innde maa?",
        "base": "What is your name?\nHow many of you are there today?\nWhat do people say when they meet each other for the first time?\nThey greet each other by saying good morning...",
        "ft": "E jaati min ngaddanii muusaa deftere nde e ɓattoore nde, tee ko o ganndo ñeeññuɗo\nE min njippini fayde to makko iwrude to kammu ngu faabaade \"dow mum\"",
    },
    {
        "prompt": "Traduis en Pulaar : Je suis fatigué.",
        "base": "Traduit par le traducteur automatique\nI am tired.\nJe suis inquiete. I'm worried.\nje ne sais pas si je peux faire l'ascension avec vous aujourd'hui.",
        "ft": "Walla: koŋngol, ina fira ɓadiiɗo,\nwaasde yande e ñalawma.\ntawi alaa baasi\n(Phase 2 en cours — traduction directe bientôt disponible)",
    },
    {
        "prompt": "Traduis en Pulaar : Bonjour, comment vas-tu ?",
        "base": "Je m'appelle Léonard et je suis un programme d'assistance virtuel. Si tu veux que j'aide à résoudre ton problème...",
        "ft": "/ Sawru? / Mbaawka ?\nPulareeɓe kala ina tawa ko heen ɗemngal mum-en waawi\ntimmude e bonnoowo\n(Phase 2 en cours — traduction directe bientôt disponible)",
    },
]


def show_result(index: int):
    r = RESULTS[index]
    return r["base"], r["ft"]


with gr.Blocks(title="PulaarLLM — Démonstration") as demo:

    gr.Markdown(
        "# PulaarLLM — Démonstration Phase 1\n"
        "**Llama 3.1 8B** fine-tuné sur **627 000 mots** de Pulaar Fouta-Toro  \n"
        "6 livres d'**Abou Sy** + Coran · Développé par **Hamath Kane**\n\n"
        "---\n"
        "### Résultats : Llama vierge vs Llama fine-tuné sur Pulaar\n"
        "Sélectionne un prompt pour voir la comparaison."
    )

    prompts = [r["prompt"] for r in RESULTS]
    prompt_selector = gr.Dropdown(
        choices=prompts,
        value=prompts[0],
        label="Prompt de test",
    )

    with gr.Row():
        with gr.Column():
            gr.Markdown("### ❌ Llama 3.1 8B vierge (sans Pulaar)")
            base_out = gr.Textbox(
                label="Réponse du modèle de base",
                lines=6,
                interactive=False,
                value=RESULTS[0]["base"],
            )
        with gr.Column():
            gr.Markdown("### ✅ Llama fine-tuné sur Pulaar (PulaarLLM v1)")
            ft_out = gr.Textbox(
                label="Réponse après fine-tuning",
                lines=6,
                interactive=False,
                value=RESULTS[0]["ft"],
            )

    def on_select(prompt):
        for r in RESULTS:
            if r["prompt"] == prompt:
                return r["base"], r["ft"]
        return "", ""

    prompt_selector.change(fn=on_select, inputs=prompt_selector, outputs=[base_out, ft_out])

    gr.Markdown(
        "---\n"
        "### Métriques d'entraînement\n"
        "| Step | Train Loss | Eval Loss |\n"
        "|------|-----------|----------|\n"
        "| 200  | 2.21      | 2.20     |\n"
        "| 400  | 1.92      | 1.96     |\n"
        "| 600  | 1.65      | **1.91** |\n\n"
        "**Corpus :** 627 000 mots · 6 livres · 3 epochs · A100 40GB · ~47 minutes\n\n"
        "---\n"
        "### Roadmap\n"
        "- ✅ **Phase 1** — Complétion de texte Pulaar (ce modèle)\n"
        "- 🔄 **Phase 2** — Traduction & dialogue en Pulaar (en cours)\n"
        "- 📱 **App Flutter** — Learn Pulaar (apprentissage mobile)\n\n"
        "**Modèle :** [kawkumputer/pulaar-llm-llama-v1](https://huggingface.co/kawkumputer/pulaar-llm-llama-v1)  \n"
        "**Traducteur NLLB :** [kawkumputer/PulaarAI](https://huggingface.co/spaces/kawkumputer/PulaarAI)"
    )

demo.launch()