Cat-v2.8XXXL-Plus

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

Overview

Maximum dataset (~3.87M entries, 1.375× XXXL). Every topic from all previous tiers — the complete Cat-v2.8 knowledge family. Uses 214 unique catgirl names. Parquet/Snappy format for efficient streaming.

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	3,870,444
Format	Chat — system / user / assistant
License	MIT
Names pool	214 unique names (incl. Nix)
Topic templates	560
File	data/train.jsonl

Entry Format

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

{
  "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

from datasets import load_dataset

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

Fine-Tuning Guide

Recommended base model

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)

# 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)

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-plus", 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	~81,000	115	General fine-tuning
Cat-v2.8HQ	~10,125	115	Fast / low-resource training
Cat-v2.8XL	~243,000	235	Deeper knowledge coverage
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.