About KAIRA

Project KAIRA is more than an AI initiative; it is a cultural manifesto. In a world dominated by generic, English-centric models, we are building Turkey's most significant open-source Large Language Model ecosystem.

Our goal is simple yet ambitious: to construct a bridge between our linguistic heritage and the future of artificial intelligence. By curating a massive 150B+ token dataset, we ensure that the Turkish language is not just represented, but mastered by the machines of tomorrow.

The Chef & The Researcher

Our founder, Umut Kökgöz, approaches AI with the discipline of a Chef. To us, data is the ingredient, and model architecture is the recipe. Just as a chef respects the source of their produce, we respect the integrity of our data sources.

Our "Kitchen" is equipped with industrial-grade tools:

• 🚀 86.4 Million Parameter VITS Models: High-fidelity Text-to-Speech synthesis with over 222 unique voices.
• 🧠 5 Billion Parameter LLMs: Custom Gemma-2 architectures pre-trained from scratch with our 104k dictionary injection.
• ♟️ Hibrit Satranç Motorları: Nöral ağları klasik arama ile birleştiren rasyonel karar alma mekanizmaları.

{
    "output_path": "/content/drive/MyDrive/KAIRA_VITS_Modeli",
    "run_name": "kaira_v35_TURKISH_MASTER",
    "project_name": "Kaira_TTS_Project",
    "run_description": "🇹🇷 Kaira VITS - Engineering Master Config (Scan Based) | Architect: [Umut Kökgöz]",
    "model": "vits",
    "epochs": 1000,
    "batch_size": 32,
    "mixed_precision": true,
    "precision": "fp16",
    "optimizer": "AdamW",
    "audio": {
        "sample_rate": 22050,
        "n_fft": 1024,
        "num_mels": 80,
        "hop_length": 256,
        "win_length": 1024,
        "fmin": 0.0,
        "fmax": null
    },
    "model_args": {
        "num_chars": 100,
        "hidden_channels": 192,
        "filter_channels": 768,
        "n_heads": 2,
        "n_layers_enc": 6,
        "kernel_size": 3,
        "p_dropout": 0.1,
        "use_sdp": true,
        "num_speakers": 223
    },
    "phonemizer": "espeak",
    "phoneme_language": "tr",
    "use_phonemes": true
}