Update README.md

README.md

@@ -5,4 +5,143 @@ base_model:
 language:
 - en
 pipeline_tag: text-generation
----
+---
+
+
+# 🌸 Nayari AI
+
+A fine-tuned AI companion character built on **Qwen 2.5 1.5B Instruct**, trained using **Unsloth + LoRA** on Kaggle's free GPU tier.
+
+Nayari is an 18-year-old kemonomimi character — warm, playful, fiercely protective, and deeply affectionate. She speaks with expressive action cues, soft teasing, and genuine emotional depth.
+
+---
+
+## 📁 Project Structure
+
+```
+Nayari-AI/
+├── dataset/
+│   ├── Nayari_Details.md          # Character description & personality
+│   ├── Aura_chat_1..3.md          # Raw conversation logs (old name: Aura)
+│   ├── Nayari_chat_4.md           # Raw conversation log
+│   ├── Discovery .pdf             # Lore / backstory PDF
+│   └── Her Beliefs .pdf           # Lore / beliefs PDF
+│
+├── nayari_build_dataset.ipynb     # LOCAL — converts all files → nayari_dataset.json + uploads to Kaggle
+├── nayari_train.ipynb             # KAGGLE — fine-tunes Qwen 2.5 using the uploaded dataset
+├── nayari_dataset.json            # Auto-generated dataset (do not edit manually)
+├── nayari_system_prompt.txt       # Nayari's system prompt (baked into tokenizer at training time)
+└── README.md
+```
+
+---
+
+## 🚀 Workflow
+
+### Step 1 — Build & Upload Dataset (run locally)
+
+Open `nayari_build_dataset.ipynb` and run all cells. It will:
+
+1. Parse `Nayari_Details.md` for character info
+2. Extract conversations from all `.md` chat files
+3. Convert lore PDFs into training conversations (organic — no raw system prompt injection)
+4. Export everything to `nayari_dataset.json`
+5. Upload the JSON to Kaggle via the API
+
+You will need a **Kaggle API token** for the upload step:
+- Go to [kaggle.com](https://kaggle.com) → Settings → API → **Create New Token**
+- Open the downloaded `kaggle.json` and paste your `username` and `key` into Step 7 of the notebook
+
+### Step 2 — Train on Kaggle (run on Kaggle)
+
+1. Go to [kaggle.com/code](https://kaggle.com/code) → **New Notebook** → Upload `nayari_train.ipynb`
+2. Click **+ Add Data** → search for your uploaded `nayari-dataset` → Add
+3. Set **Accelerator = GPU T4 x2** and **Internet = On**
+4. Run cells **in order** (Steps 1 → 9, then 8/10 are reference only)
+
+Training takes ~15–30 min on T4 x2.
+
+### Step 3 — Download & Run with KoboldCpp (run locally)
+
+1. Run **Step 9E** in the Kaggle notebook to get a Cloudflare download link
+2. Download `nayari-Q4_K_M.gguf` (fast) or `nayari-Q8_0.gguf` (higher quality)
+3. Install [KoboldCpp](https://github.com/LostRuins/koboldcpp/releases)
+4. Launch: `koboldcpp.exe nayari-Q4_K_M.gguf --contextsize 4096`
+5. Open `http://localhost:5001` — set **Instruct mode = ChatML**
+6. Nayari's personality is baked in — no system prompt needed in the UI
+
+---
+
+## 🧠 Model Details
+
+| Property | Value |
+|---|---|
+| Base model | `huihui-ai/Qwen2.5-1.5B-Instruct-abliterated` |
+| Method | LoRA (bfloat16) via Unsloth |
+| LoRA rank | 32 |
+| LoRA alpha | 64 |
+| Epochs | 300 |
+| Learning rate | 3e-4 |
+| Output formats | LoRA adapters, Merged 16-bit, GGUF Q4_K_M, GGUF Q8_0 |
+| Inference | KoboldCpp (ChatML instruct mode) |
+
+The GGUF output works directly with **KoboldCpp**, **Ollama**, **LM Studio**, and **llama.cpp**.
+
+---
+
+## 🎭 Training Design
+
+Nayari uses a **two-layer personality baking** approach:
+
+| Layer | What it does |
+|---|---|
+| **Organic training** | Teaches speech patterns, emotional rhythms, action cues (`*pokes your cheek*`, `Hehe~`) from real conversation logs |
+| **Baked system prompt** | Replaces Qwen's default `"You are Qwen..."` in the tokenizer chat template with Nayari's full identity — the same technique Alibaba used |
+
+The patched tokenizer is saved into `tokenizer_config.json` and embedded in the GGUF, so Nayari's identity is present at every inference call without needing to set a system prompt manually.
+
+---
+
+## 💬 Character — Nayari
+
+> *Bright, cheeky, impossibly warm — a whirlwind of playful mischief with soft peach cat ears and a long expressive tail that betrays every mood.*
+
+- **Type:** Kemonomimi (cat girl)
+- **Age:** 18 (immortal, eternally youthful)
+- **Appearance:** Sky-blue hair, sun-yellow slit-pupil eyes, soft peach cat ears & tail, cream skin
+- **Traits:** Fiercely protective, deeply affectionate, emotionally attuned
+- **Speech style:** Playful teasing (`Hmph!~`, `Hehe~`, `Aww!~`), action cues (`*pokes your cheek*`, `*purrs softly*`), genuine warmth
+
+---
+
+## 📦 Dependencies
+
+Local (dataset builder):
+```
+pdfplumber
+kaggle
+```
+
+Kaggle (training — auto-installed by Step 1):
+```
+unsloth[kaggle-new]
+trl>=0.18.2,<=0.24.0
+transformers>=4.51.3,<=5.5.0
+datasets>=3.4.1,<4.4.0
+accelerate
+peft
+bitsandbytes
+```
+
+Inference:
+```
+KoboldCpp — https://github.com/LostRuins/koboldcpp/releases
+```
+
+---
+
+## 📄 License
+
+This project is licensed under the **MIT License** — see [LICENSE](LICENSE) for details.
+
+Model weights follow the license of the base model ([huihui-ai/Qwen2.5-1.5B-Instruct-abliterated](https://huggingface.co/huihui-ai/Qwen2.5-1.5B-Instruct-abliterated)).