Instructions to use Crossie/Nayari with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Crossie/Nayari with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="Crossie/Nayari")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Crossie/Nayari")
model = AutoModelForCausalLM.from_pretrained("Crossie/Nayari")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use Crossie/Nayari with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Crossie/Nayari"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Crossie/Nayari",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Crossie/Nayari

SGLang

How to use Crossie/Nayari with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Crossie/Nayari" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Crossie/Nayari",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Crossie/Nayari" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Crossie/Nayari",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Unsloth Studio new

How to use Crossie/Nayari with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for Crossie/Nayari to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for Crossie/Nayari to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for Crossie/Nayari to start chatting

Load model with FastModel

pip install unsloth
from unsloth import FastModel
model, tokenizer = FastModel.from_pretrained(
    model_name="Crossie/Nayari",
    max_seq_length=2048,
)

Docker Model Runner
How to use Crossie/Nayari with Docker Model Runner:
```
docker model run hf.co/Crossie/Nayari
```

Crossie commited on 11 days ago

Commit

c80ea21

verified ·

1 Parent(s): 1639328

Update README.md

Browse files

Files changed (1) hide show

README.md +38 -65

README.md CHANGED Viewed

@@ -13,6 +13,8 @@ A fine-tuned AI companion character built on **Qwen 2.5 1.5B Instruct**, trained
 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
@@ -44,53 +46,56 @@ 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**.
-Training takes ~15–30 min on T4 x2.
-### Step 3 — Export & Download (run on Kaggle)
-1. After training, open `nayari_export.ipynb` in your Kaggle environment (or upload it to a new notebook with the same workspace context).
-2. Run the cells to generate LoRA adapters, merged 16-bit, and GGUF outputs.
-3. Use the **Cloudflare Tunnel** cell to get direct HTTP download links for the generated `.gguf` files.
-4. Download `nayari-Q4_K_M.gguf` (fast) or `nayari-Q8_0.gguf` (higher quality).
-### Step 4 — Run with KoboldCpp (run locally)
-1. Install [KoboldCpp](https://github.com/LostRuins/koboldcpp/releases)
-2. Launch: `koboldcpp.exe nayari-Q4_K_M.gguf --contextsize 4096`
-3. Open `http://localhost:5001` — set **Instruct mode = ChatML**
-4. 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**.
 ---
@@ -98,12 +103,8 @@ The GGUF output works directly with **KoboldCpp**, **Ollama**, **LM Studio**, an
 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.
 ---
@@ -115,38 +116,10 @@ The patched tokenizer is saved into `tokenizer_config.json` and embedded in the
 - **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)).

 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.
+**🔗 Model Weights:** [https://huggingface.co/Crossie/Nayari](https://huggingface.co/Crossie/Nayari)
 ---
 ## 📁 Project Structure
 1. Parse `Nayari_Details.md` for character info
 2. Extract conversations from all `.md` chat files
+3. Convert lore PDFs into training conversations
 4. Export everything to `nayari_dataset.json`
 5. Upload the JSON to Kaggle via the API
 ### Step 2 — Train on Kaggle (run on Kaggle)
+1. Go to [kaggle.com/code](https://kaggle.com) → **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**.
+### Step 3 — Export & Download
+1. After training, use `nayari_export.ipynb` to generate LoRA adapters and GGUF outputs.
+2. You can download the files directly or push them to Hugging Face.
+3. Pre-compiled weights are available here: [Crossie/Nayari](https://huggingface.co/Crossie/Nayari)
+---
+## 💻 How to Run Nayari
+Nayari is exported in **GGUF** and **Hugging Face** formats, making her compatible with almost any modern LLM runner.
+### Option A: Desktop Apps (Easiest)
+Download the `.gguf` file from the [Hugging Face repo](https://huggingface.co/Crossie/Nayari) and load it into:
+*   **LM Studio:** Search for the local file and load it. Use the "ChatML" preset.
+*   **KoboldCpp:** Excellent for roleplay. Launch and set **Instruct Mode = ChatML**.
+*   **Jan.ai / AnythingLLM:** Standard GGUF support.
+### Option B: Command Line / Servers
+*   **Ollama:** Create a `Modelfile` pointing to the GGUF and run `ollama create nayari -f Modelfile`.
+*   **Llama.cpp:** Run via `./llama-cli -m nayari-Q4_K_M.gguf -p "ChatML"`.
+### Important Settings
+*   **Instruct Mode / Prompt Template:** ChatML
+*   **System Prompt:** Not required (Nayari’s identity is baked into the model's tokenizer).
+*   **Context Length:** 4096 (or higher if your hardware allows).
 ---
 ## 🧠 Model Details
+| Property              | Value                                                                  |
+| -----------------------| ------------------------------------------------------------------------|
+| **Base model**        | `huihui-ai/Qwen2.5-1.5B-Instruct-abliterated`                          |
+| **Weights**           | [huggingface.co/Crossie/Nayari](https://huggingface.co/Crossie/Nayari) |
+| **Method**            | LoRA (bfloat16) via Unsloth                                            |
+| **LoRA rank / alpha** | 32 / 64                                                                |
+| **Epochs**            | 300                                                                    |
+| **Output formats**    | GGUF (Q4, Q8), FP16, LoRA Adapters                                     |
 ---
 Nayari uses a **two-layer personality baking** approach:
+1.  **Organic Training:** Teaches speech patterns, emotional rhythms, and action cues (`*pokes your cheek*`, `Hehe~`) from real conversation logs.
+2.  **Baked System Prompt:** Nayari's full identity is embedded directly into the tokenizer's chat template. This means the model "knows" who she is from the first token without the user needing to provide a long system description in the UI.
 ---
 - **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~`), action cues (`*purrs softly*`), genuine warmth
 ---
 ## 📄 License
+This project is licensed under the **MIT License**. Model weights follow the license of the base model ([Qwen 2.5](https://huggingface.co/collections/Qwen/qwen25-66e81a6663533ad4ab30046b)).