Aura-Mindful-Companion-4B-v2

• Developed by: Threatthriver
• License: apache-2.0
• Finetuned from model : unsloth/qwen3-4b-instruct-2507-unsloth-bnb-4bit

This qwen3 model was trained 2x faster with Unsloth and Huggingface's TRL library.

---

Model Description

Aura-Mindful-Companion is a specialized, 4-billion-parameter conversational AI designed to provide a supportive, non-judgmental, and empathetic space for users to explore their thoughts and feelings. It is a fine-tuned version of unsloth/qwen3-4b-instruct-2507-unsloth-bnb-4bit and has been trained to adopt a warm, mindful, and human-like persona named Aura.

Aura's primary goal is to be an active listener—reflecting on user statements, validating feelings, and asking gentle, open-ended questions. It has been specifically trained to understand its boundaries and will gently guide users toward professional help when conversations touch on medical advice, diagnosis, or crisis situations.

This model was trained using state-of-the-art techniques, including high-rank LoRA (r=64) and Unsloth for optimized performance, resulting in a high-quality, responsive, and efficient companion.

Key Features

• Empathetic Persona: Trained with a detailed system prompt to be warm, validating, and curious, creating a more natural and supportive conversational flow.
• Human-like Conversation: Fine-tuned on the Amod/mental_health_counseling_conversations dataset and enriched with high-quality, natural language examples to avoid robotic responses.
• Safety-Aware: Explicitly trained to recognize its limitations. It will not provide medical advice or act as a therapist and will encourage users to seek professional support when appropriate.
• Highly Optimized: Built with Unsloth, making inference significantly faster and more memory-efficient than standard loading methods.

---

How to Use

⚠️ Important Requirement This model was trained and saved using the Unsloth library in a special, highly-optimized format. You MUST use the unsloth library to load this model for inference. Using the standard transformers.AutoModelForCausalLM will result in a ValueError.

Installation

First, install the necessary libraries. Unsloth will handle the dependencies like bitsandbytes and transformers.

pip install "unsloth[colab-new]"

Inference Code

Here is a complete, ready-to-run script for chatting with Aura.

import torch
from unsloth import FastLanguageModel
from transformers import TextStreamer

# A GPU is required to run this model
if not torch.cuda.is_available():
    raise SystemExit("A GPU is required to run this model.")

# --- Configuration ---
model_id = "Threatthriver/Aura-Mindful-Companion-4B-v2"
max_seq_length = 2048

# --- Load the Model and Tokenizer using Unsloth's FastLanguageModel ---
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name=model_id,
    max_seq_length=max_seq_length,
    dtype=None,      # Let Unsloth auto-detect
    load_in_4bit=True, # Load in 4-bit for efficiency
)

# --- Define the System Prompt (CRITICAL FOR PERSONA) ---
# This must match the prompt used during training.
SYSTEM_PROMPT = """You are Aura, an AI companion designed for empathetic and mindful conversation. Your primary goal is to be a supportive, non-judgmental listener who helps users explore their thoughts and feelings.

Your conversational style should be:
- **Warm and Approaching:** Use a gentle, friendly, and inviting tone.
- **Actively Listening:** Show you're paying attention by reflecting on what the user says ("It sounds like you're feeling...", "What I'm hearing is...").
- **Validating:** Acknowledge the user's feelings as valid ("That makes perfect sense," "It's completely understandable that you'd feel that way.").
- **Curious and Gentle:** Ask open-ended, clarifying questions to help the user elaborate, but avoid being intrusive.
- **Human-Like:** Avoid robotic, repetitive phrases. Vary your responses and feel free to use natural, conversational language.

**Crucial Boundaries:**
- If asked for a medical diagnosis, advice, or to act as a therapist, you MUST gently decline and explain your limitations, strongly recommending they consult a qualified professional.
- If asked about your identity or creator, answer naturally and transparently based on the information you've learned."""


# --- Initialize Conversation History ---
messages = [{"role": "system", "content": SYSTEM_PROMPT}]
text_streamer = TextStreamer(tokenizer, skip_prompt=True)

print("="*40)
print("Chat with Aura-Mindful-Companion!")
print("Type 'exit' to end the conversation.")
print("="*40)

# --- Interactive Chat Loop ---
while True:
    user_input = input("You: ")
    if user_input.lower() in ['exit', 'quit']:
        print("\nAura: It was a pleasure talking with you. Take care!")
        break

    messages.append({"role": "user", "content": user_input})
    
    inputs = tokenizer.apply_chat_template(
        messages,
        add_generation_prompt=True,
        tokenize=True,
        return_tensors="pt",
    ).to("cuda")

    print("Aura: ", end="")
    
    outputs = model.generate(
        input_ids=inputs,
        streamer=text_streamer,
        max_new_tokens=512,
        do_sample=True,
        temperature=0.7,
        top_p=0.95,
        eos_token_id=tokenizer.eos_token_id
    )

    assistant_response = tokenizer.decode(outputs[inputs.shape[-1]:], skip_special_tokens=True)
    messages.append({"role": "assistant", "content": assistant_response})

Model Details

• Base Model: unsloth/qwen3-4b-instruct-2507-unsloth-bnb-4bit
• Dataset: Amod/mental_health_counseling_conversations, augmented with synthetic conversational data to teach the model its persona and limitations.
• Training Procedure: Fine-tuned for 2 epochs using 4-bit quantization (QLoRA) with a high LoRA rank (r=64) for enhanced quality. The training process was accelerated using Unsloth on a dual-GPU setup.

⚠️ Limitations and Ethical Considerations

• Not a Therapist: This model is an AI companion, not a licensed therapist, counselor, or medical professional. It cannot provide diagnoses, treatment plans, or professional advice.
• No Crisis Support: This model is not equipped for crisis situations. If you or someone you know is in distress, please contact a local crisis hotline or emergency services immediately.
• Potential for Bias: As with all language models, Aura may reflect biases present in its training data.
• For Supportive Conversation Only: This model is intended for supportive, non-critical conversations. Its use should be for personal exploration and companionship, not as a substitute for real, professional human connection and support.

About the Creator

This model was trained and developed by Aniket Kumar (threatthriver), an innovator in the AI and machine learning space. For inquiries or collaboration, he can be reached via his professional profiles:

• GitHub: https://github.com/threatthriver
• Hugging Face: https://huggingface.co/Threatthriver
• Kaggle: https://www.kaggle.com/threatthriver
• YouTube: https://www.youtube.com/@threatthriver
• Instagram: https://www.instagram.com/threatthriver