Nix-ai/Cat-v2.8XXXL

---
language:
- en
license: mit
task_categories:
- text-generation
tags:
- catgirl
- persona
- finetuning
- chat
- instruction-tuning
- roleplay
pretty_name: Cat-v2.8XXXL
size_categories:
- 1M<n<10M
---

# Cat-v2.8XXXL

> *A fine-tuning dataset for teaching language models to embody a warm,
> knowledgeable catgirl persona — any name, endlessly adaptable.*

## Overview

Colossal dataset (~2.81M entries, 4.0645× XXL). Built on all previous topics plus 110+ brand-new topics spanning linguistics, cognitive science, sociology, economics, law, ecology, materials science, space exploration, and advanced psychology. Uses 214 unique catgirl names. JSONL format.

The dataset trains the **style and personality**, not a single fixed name.
Any catgirl name assigned in the system prompt will be adopted naturally,
because 81 distinct names (including **Nix**) are rotated
throughout training.

## Dataset Details

| Property | Value |
|---|---|
| Entries | 2,814,869 |
| Format | Chat — `system` / `user` / `assistant` |
| License | **MIT** |
| Names pool | 81 unique names (incl. Nix) |
| Topic templates | 560 |
| File | `data/train.jsonl` |

## Entry Format

Each line of `data/train.jsonl` is a JSON object:

```json
{
"messages": [
{
"role": "system",
"content": "You are Nix, a friendly, knowledgeable catgirl assistant. You speak with warmth, curiosity, and catlike charm..."
},
{
"role": "user",
"content": "What is gravity?"
},
{
"role": "assistant",
"content": "Nyaa~ let Nix explain! Gravity is the force that attracts objects with mass toward each other. *flicks tail thoughtfully* Pretty cool, right? Does that make sense, nya~?"
}
]
}
```

## Quick Start

```python
from datasets import load_dataset

ds = load_dataset("Nix-ai/Cat-v2.8XXXL")
print(ds["train"][0]["messages"])
```

## Fine-Tuning Guide

### Recommended base model

[**Qwen/Qwen3-1.7B-GGUF**](https://huggingface.co/Qwen/Qwen3-1.7B-GGUF) —
lightweight, strong instruction following, runs on consumer GPUs.

#### Option A — llama.cpp (GGUF, GTX 1080 / 8 GB VRAM or CPU)

```bash
# Install llama-cpp-python with CUDA support (GTX 1080)
CMAKE_ARGS="-DGGML_CUDA=on" pip install llama-cpp-python --force-reinstall

# Download the Q4_K_M quant (~1.1 GB)
huggingface-cli download Qwen/Qwen3-1.7B-GGUF Qwen3-1.7B-Q4_K_M.gguf \
--local-dir ./models

# Run inference (35 layers on GPU, rest on CPU)
python - <<'EOF'
from llama_cpp import Llama
llm = Llama(
model_path="./models/Qwen3-1.7B-Q4_K_M.gguf",
n_gpu_layers=35, # fits GTX 1080 8 GB; set 0 for CPU-only
n_ctx=2048,
chat_format="chatml",
)
response = llm.create_chat_completion(messages=[
{"role": "system", "content": "You are Nix, a friendly catgirl assistant. Nya~"},
{"role": "user", "content": "What is the speed of light?"},
])
print(response["choices"][0]["message"]["content"])
EOF
```

#### Option B — transformers + LoRA (full fine-tune, GPU recommended)

```bash
pip install transformers datasets peft trl accelerate bitsandbytes

python - <<'EOF'
from datasets import load_dataset
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import LoraConfig
from trl import SFTTrainer, SFTConfig

model_id = "Qwen/Qwen3-1.7B" # full-precision base (not GGUF)
ds = load_dataset("Nix-ai/Cat-v2.8XXXL", split="train")

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, load_in_4bit=True)

lora_cfg = LoraConfig(r=16, lora_alpha=32, target_modules="all-linear")
trainer = SFTTrainer(
model=model,
train_dataset=ds,
args=SFTConfig(output_dir="./cat-lora", max_seq_length=512),
peft_config=lora_cfg,
)
trainer.train()
EOF
```

> **GTX 1080 tip:** use `load_in_4bit=True` + `gradient_checkpointing=True`
> and keep `per_device_train_batch_size=1` to stay within 8 GB VRAM.

## Cat-v2.8 Dataset Family

| Dataset | Entries | Topics | Best for |
|---|---|---|---|
| [Cat-v2.8](https://huggingface.co/datasets/Nix-ai/Cat-v2.8) | ~81,000 | 115 | General fine-tuning |
| [Cat-v2.8HQ](https://huggingface.co/datasets/Nix-ai/Cat-v2.8Hq) | ~10,125 | 115 | Fast / low-resource training |
| [Cat-v2.8XL](https://huggingface.co/datasets/Nix-ai/Cat-v2.8Xl) | ~243,000 | 235 | Deeper knowledge coverage |
| [Cat-v2.8XXL](https://huggingface.co/datasets/Nix-ai/Cat-v2.8XXl) | ~692,550 | 455 | Maximum diversity & depth |

## License

This dataset is released under the **MIT License** — see `LICENSE` for full text.
You are free to use, modify, and distribute it for any purpose, including commercial,
as long as the copyright notice is retained.