RikkaBotan/LFM2-1.2B-Cute-Friendly-Finetune-JP-GGUF

🌸 LFM2–Friendly Japanese Fine-Tuned Model

A warm, approachable, and soft-spoken conversational AI

This repository provides a fine-tuned version of LFM2 (Liquid Foundation Model v2) designed to deliver gentle, friendly, and natural Japanese conversations. The model has been trained to speak in a soft, feminine, and approachable style, similar to a kind and caring girl.

✨ Overview

This model adapts LFM2 with additional fine-tuning on a curated Japanese dataset to emphasize:

• 🌷 Warm and approachable tone
• 💕 Soft, gentle, girl-like speaking style
• 🗣️ Natural and smooth Japanese dialog
• 🤝 Supportive and friendly communication

The goal is to create an AI that feels relaxing to talk to, while still maintaining the strengths of LFM2 such as stability, reasoning ability, and responsiveness.

🎯 Fine-Tuning Details

• Base Model: LFM2
• Language: Japanese
• Style Objective: Soft feminine tone, kind responses, polite casual phrasing
• Training Strategy: Supervised fine-tuning (SFT) on custom conversational data designed to reinforce emotional warmth and friendliness

💬 Example Characteristics

• Uses gentle sentence endings (〜だよ, 〜かな?, 〜だね) depending on context
• Encourages positive and comfortable interactions
• Avoids sharp or overly formal phrasing unless context requires it
• Maintains clarity while sounding cute and approachable

🔧 Usage

You can load the model like any standard LFM2-compatible checkpoint:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "RikkaBotan/LFM2-350M-Cute-Friendly-Finetune-JP"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

input_text = "今日はちょっと疲れちゃった…"
outputs = model.generate(tokenizer.encode(input_text, return_tensors="pt"))

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

📌 Intended Applications

• ✨ Casual Japanese conversation
• ✨ Companion-like dialog
• ✨ Emotional support–style interactions
• ✨ Storytelling and character-based responses

⚠️ Note: While the model is fine-tuned for a gentle “girl-like” speaking style, it does not represent a real person and should not be used for inappropriate or harmful purposes.