fitness-buddy-1.0

Fitness Buddy 1.0

This is a fine-tuned Llama-3.2-1B model using LoRA for fitness Q&A. It provides accurate, concise answers to questions about workouts, nutrition, and fitness routines based on custom training data.

Base Model

• Base Model: Soorya03/Llama-3.2-1B-FitnessAssistant

Setup Instructions

Prerequisites

• Python 3.8+
• Install required packages:

  pip install transformers peft torch
  

Download and Setup

1. Clone or download this repository.
2. Ensure you have access to the model files (adapter_model.safetensors, tokenizer files, etc.) in this folder.
3. The model is hosted on Hugging Face at sarveshcore/fitness-buddy-1.0.

Usage

Loading the Model

Use the following Python script to load and test the model. This merges the base model with your fine-tuned LoRA adapter.

from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
from peft import PeftModel

# Load the base model and your fine-tuned LoRA adapter
base_model_name = "Soorya03/Llama-3.2-1B-FitnessAssistant"
base_model = AutoModelForCausalLM.from_pretrained(base_model_name)
model = PeftModel.from_pretrained(base_model, "sarveshcore/fitness-buddy-1.0")
tokenizer = AutoTokenizer.from_pretrained("sarveshcore/fitness-buddy-1.0")

# Create the pipeline
pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)

# Test with a sample question
test_question = "Question: How often should I do PPL?\n\nAnswer:"
output = pipe(
    test_question,
    max_new_tokens=500,  # Increase for longer responses
    do_sample=True,
    temperature=0.7,     # Slightly higher for more detail
    top_p=0.95,
    repetition_penalty=1.1  # Lower penalty
)

# Print the response
generated_text = output[0]['generated_text']
answer = generated_text.replace(test_question, "").strip()
print("Fitness Assistant:", answer)

Running the Test

• Save the code above as test_model.py.
• Run: python test_model.py
• Expected output: A detailed answer based on fitness knowledge (e.g., "3–6 days/week, e.g., PPL rest repeat. Hits muscles 2x/week for growth.").

Integration into Chatbot

To use in a chatbot (like test_granite.py):

from transformers import pipeline

pipe = pipeline("text-generation", model="sarveshcore/fitness-buddy-1.0")

# In your loop
user_question = input("You: ")
formatted_prompt = f"Question: {user_question}\n\nAnswer:"
output = pipe(formatted_prompt, max_new_tokens=300, temperature=0.5)
answer = output[0]['generated_text'].replace(formatted_prompt, "").strip()
print("Fitness Assistant:", answer)

Model Evaluation

We performed a preliminary automated benchmark to test the model on three core capabilities: Safety, Factual Knowledge, and Routine Generation.

Current Limitations

• Safety: The model provides general recovery advice (rest, ice) but does not consistently trigger a "medical disclaimer" for injury queries.
• Structured Output: While the model suggests valid exercises, it currently struggles to adhere to strict formatting (e.g., specific set/rep counts) even when prompted with examples.

🚀 Future Roadmap: Project "Hybrid Athlete"

Our goal is to evolve fitness-buddy from a simple Q&A bot into a comprehensive "Fitness Instructor" capable of handling complex biomechanics and nutrition.

Due to local hardware constraints, we are opening this project to contributors to help implement a Hybrid Athlete architecture:

1. 🧠 Knowledge Expansion (RAG)

• The Goal: Implement Retrieval-Augmented Generation (RAG) to connect the model to verified external sources (e.g., anatomy databases, nutritional journals).
• The Benefit: Allows the model to answer "anything" by reading reliable data in real-time, rather than relying solely on memorized parameters.

2. 🧬 "Genetics" Upgrade (Larger Base Models)

• The Goal: Port the fine-tuning dataset to larger base models (e.g., Llama-3-8B or Mistral-7B).
• The Benefit: Larger models have significantly better reasoning capabilities and are more resistant to "catastrophic forgetting" (losing old facts while learning new routines).

3. 🏋️‍♂️ Volume Phase (Synthetic Data)

• The Goal: Create a pipeline to generate high-quality synthetic training data (10k+ examples).
• The Benefit: Exposes the model to a wider variety of workout structures and edge-case safety scenarios that cannot be written by hand.

Training Data

Trained on custom fitness Q&A pairs covering bulking, cutting, workouts (PPL, upper/lower, full body), rest days, and more.

License

MIT

Tags

• fitness
• lora
• fine-tuned
• llama
• datasets: custom