Jahirrrr/ur-own-gf

💖 UR OWN GIRLFRIEND!

UR OWN GF is a high-fidelity roleplay model finetuned on the Ministral-3B-Instruct base model.

It has been specifically trained to simulate a Girlfriend (GF) persona, focusing on natural conversation, emotional intelligence, and engaging roleplay interactions.

This model avoids the "robotic" feel of standard assistants and is optimized for casual, affectionate, and dynamic roleplay.

🚀 Model Highlights

This model was trained using Unsloth on an NVIDIA H100 (80GB) GPU with maximum quality settings to ensure the model captures the finest nuances of the dataset:

• Base Model: Ministral 3B (State-of-the-art small model, outperforms many 7B models).
• Precision: Native BFloat16 (No quantization during training for maximum intelligence).
• LoRA Rank: 32 Lower rank chosen to learn style without memorizing specific entities.
• Context Window: 4096 Tokens.
• Dataset: Jahirrrr/gf-conversation.

📦 Available GGUF Quantizations

This repository contains GGUF files compatible with LM Studio, Ollama, KoboldCPP, and llama.cpp.

Filename	Quantization	Size	Description
ur-own-gf-f16.gguf	F16	~6.5 GB	Best Quality. Uncompressed weights. Recommended if you have >8GB VRAM.
ur-own-gf-q8_0.gguf	Q8_0	~3.5 GB	Near-lossless quality. Very fast processing. Recommended for 6GB+ VRAM.
ur-own-gf-q5_k_m.gguf	Q5_K_M	~2.4 GB	High accuracy, balanced size.
ur-own-gf-q4_k_m.gguf	Q4_K_M	~2.0 GB	Recommended for Mobile/Low VRAM. Good balance of speed and smarts.

🗣️ Chat Template (Mistral)

This model uses the standard Mistral chat template.

💻 How to Use

Option 1: LM Studio / Ollama (Easiest)

1. Download the .gguf file of your choice (e.g., q4_k_m or q8_0).
2. Load it into LM Studio.
3. Set the System Prompt (Optional) to something like: "You are a loving and caring girlfriend."
4. Start chatting!

Option 2: Python (Unsloth/Transformers)

You can run the model (GGUF or LoRA) directly in Python using Unsloth.

from unsloth import FastLanguageModel

# 1. Load the model
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "Jahirrrr/ur-own-gf",
    max_seq_length = 4096,
    load_in_4bit = True, # Set False for F16
)
FastLanguageModel.for_inference(model)

# 2. Prepare Message
messages = [
    {"role": "user", "content": "Hi, babe!"},
]

# 3. Apply Chat Template
inputs = tokenizer.apply_chat_template(
    messages, 
    tokenize=True, 
    add_generation_prompt=True, 
    return_tensors="pt"
).to("cuda")

# 4. Generate Response
outputs = model.generate(
    input_ids=inputs, 
    max_new_tokens=128, 
    temperature=0.7,
    top_p=0.9
)

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

🛠️ Training Details

This model was trained using Unsloth on high-performance hardware. The training parameters were specifically tuned to balance style adaptation (learning the "girlfriend" persona) while minimizing knowledge overfitting.

⚙️ Infrastructure & Environment

Parameter	Value	Description
Hardware	NVIDIA H100 (80GB)	High-end datacenter GPU for maximum speed and precision.
Library	Unsloth	Accelerated fine-tuning library (2x faster).
Base Model	Ministral-3B-Instruct	State-of-the-art small language model (2512 version).
Dataset	Jahirrrr/gf-conversation	Conversational roleplay dataset.
Precision	BFloat16 (BF16)	Native 16-bit precision (No 4-bit quantization used during training).

🧠 LoRA Configuration

Parameter	Value	Notes
Rank (r)	32	Lower rank chosen to learn style without memorizing specific entities.
LoRA Alpha	64	Scaling factor (Standard 2x Rank).
LoRA Dropout	0.05	Added to increase model flexibility and prevent overfitting.
Target Modules	All Linear Layers	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj.
Bias	None	Optimized for efficiency.

📊 Hyperparameters

Parameter	Value	Context
Context Window	4096 Tokens	Sufficient for medium-length chat history.
Learning Rate	2e-5	Standard SFT learning rate.
Batch Size	32	Global batch size (per_device: 16 × accumulation: 2).
Max Steps	60	Short training duration (~1.5 epochs) to avoid rote memorization of dataset names.
Optimizer	AdamW 8-bit	Memory-efficient optimizer.
Scheduler	Cosine	Smooth convergence.
Weight Decay	0.01	Regularization parameter.
Training Method	Responses Only	Masking applied on user prompts (Model learns to answer, not to ask).

---

Model trained & quantized by Jahirrrr