---
language: en
license: apache-2.0
tags:
- llama-3.2
- fine-tuning
- meditation
- guided-meditation
- wellness
- text-generation
base_model: meta-llama/Meta-Llama-3.2-3B-Instruct
datasets:
- AlbertoB12/GuidedMeditations1
---

# Meditation Guide (Llama 3.2 - 3B)

This is a fine-tuned version of `meta-llama/Meta-Llama-3.2-3B-Instruct`, specifically adapted to generate guided meditation scripts. The model was trained on the `AlbertoB12/GuidedMeditations1` dataset, a collection of diverse guided meditation texts.

The goal of this project is to provide a specialized AI tool for creating content in the wellness and mindfulness space. It can generate complete meditation scripts based on a simple prompt, focusing on themes like relaxation, anxiety relief, focus, and gratitude.

## Model Description

- **Base Model**: `meta-llama/Meta-Llama-3.2-3B-Instruct`
- **Language**: English (en)
- **Task**: Text Generation, Guided Meditation Scripting
- **Trained on:** [https://huggingface.co/datasets/AlbertoB12/GuidedMeditations1](https://huggingface.co/datasets/AlbertoB12/GuidedMeditations1)

The model excels at adopting a calm, encouraging, and guiding tone suitable for meditation. It understands instructions related to pacing, focus points (e.g., breath, body sensations), and common meditation themes.

## Intended Uses & Limitations

### Intended Uses

This model is designed for:
-   **Content Creation**: Generating scripts for wellness apps, YouTube channels, or personal mindfulness practice.
-   **Personalization**: Creating custom meditation scripts tailored to specific needs (e.g., "a 5-minute meditation for morning focus").
-   **Creative Assistance**: A tool for mindfulness teachers and practitioners to brainstorm and develop new meditation content.

> **Disclaimer:** This model is for informational and creative purposes only. The content it generates is **not** a substitute for professional medical or psychological advice, diagnosis, or treatment.

### Limitations

-   **Narrow Domain**: The model is highly specialized. It may not perform well on topics outside of meditation, mindfulness, and general wellness.
-   **Potential for Hallucination**: Like all LLMs, it may occasionally generate text that is nonsensical or not perfectly aligned with the prompt.
-   **Bias**: The model's output will reflect the styles and potential biases present in the `GuidedMeditations1` dataset.

## How to Use

To use this model, ensure you have accepted the terms of use for Llama 3.2 on the `meta-llama/Meta-Llama-3.2-8B-Instruct` model page. The model should be used with the Llama 3.2 chat template.

```python
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
import os

# --- Configuration ---
# Set your Hugging Face token (if the model is private or requires authentication)
# For HF Spaces, set this as a secret named HF_TOKEN
hf_token = os.getenv("HF_TOKEN") 
model_id = "AlbertoB12/Llama-3.2-3B-Instruct-MeditationGuide"

# --- Load Tokenizer and Model ---
tokenizer = AutoTokenizer.from_pretrained(model_id, token=hf_token, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    token=hf_token,
    trust_remote_code=True
)
model.eval()

# --- Prepare the Prompt ---
# Use the official chat template for Llama 3.2
messages = [
    {
        "role": "system",
        "content": "You are a helpful meditation guide. Your purpose is to generate calm, soothing, and effective guided meditation scripts based on the user's request."
    },
    {
        "role": "user",
        "content": "Write a 5-minute guided meditation script focused on releasing anxiety."
    },
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

# --- Generate the Response ---
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=1024,
    do_sample=True,
    temperature=0.7,
    top_p=0.95,
    eos_token_id=tokenizer.eos_token_id
)

# --- Decode and Print ---
response = tokenizer.decode(outputs[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True)
print(response)
```