Hugging Face's logo

a-albiol
/
AthenAI

Safetensors
t5
v2
Model card Files
xet
Community

AthenAI - AI-Powered Personalized Workout Generator

AthenAI is a fine-tuned FLAN-T5-base model designed to generate personalized workout routines based on user context, training goals, motivation, and special health situations. The model creates structured workout plans with multiple exercise blocks, duration estimates, and detailed instructions in JSON format.

Model Details

Model Description

AthenAI takes user context as input (training phase, motivation, special situations) and generates comprehensive workout plans tailored to individual needs. The model was trained on synthetic workout data derived from comprehensive exercise databases and can handle various fitness scenarios from weight loss to injury recovery.

  • Developed by: a-albiol
  • Model type: Text-to-Text Generation (Sequence-to-Sequence)
  • Language(s) (NLP): English (primary), Spanish (secondary)
  • License: MIT
  • Finetuned from model: google/flan-t5-base

Model Sources

Uses

Direct Use

AthenAI is designed for direct use in generating personalized workout routines. Users can input their training context and receive structured workout plans immediately. The model is particularly useful for:

  • Personal Fitness Applications: Generate daily workout routines
  • Fitness Apps: Provide adaptive exercise recommendations
  • Gym Management Systems: Create member-specific workout plans
  • Health & Wellness Platforms: Offer personalized fitness guidance

Downstream Use

The model can be integrated into larger fitness ecosystems:

  • Mobile Fitness Apps: Backend workout generation service
  • Personal Training Software: Assist trainers with plan creation
  • Rehabilitation Systems: Generate recovery-focused exercise routines
  • Corporate Wellness Programs: Provide employee fitness plans

Out-of-Scope Use

AthenAI should not be used for:

  • Medical diagnosis or treatment recommendations
  • Professional medical or physiotherapy advice
  • Unsupervised use for individuals with serious cardiovascular conditions
  • Replacement for professional fitness consultation
  • Legal or liability-bearing fitness recommendations

Bias, Risks, and Limitations

Known Limitations

  1. Exercise Database Scope: Primarily trained on gym-based exercises, limited outdoor/home alternatives
  2. Equipment Assumptions: May suggest exercises without considering equipment availability
  3. Medical Expertise: Cannot replace professional medical or fitness consultation
  4. Cultural Context: Training data may reflect Western fitness practices
  5. Language Limitations: Optimized for English with limited Spanish support

Potential Biases

  • Dataset Bias: Limited to exercises popular in online fitness communities
  • Demographic Bias: May not adequately represent all age groups, fitness levels, or cultural backgrounds
  • Equipment Bias: Assumes access to standard gym equipment
  • Ability Bias: May not fully accommodate all physical limitations or disabilities

Recommendations

Users should be aware that AthenAI:

  • Provides general fitness guidance, not medical advice
  • Should be used alongside professional consultation for special health conditions
  • May require human review for individuals with physical limitations
  • Works best when combined with proper fitness supervision

How to Get Started with the Model

Installation 

pip install transformers torch

Basic Usage 

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

# Load model and tokenizer
model_name = "a-albiol/AthenAI"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)

# Define user context
context = {
    "training_phase": "weight_loss",
    "motivation": "wellbeing", 
    "special_situation": "none"
}

# Generate workout
input_text = f"Generate a workout for: {context}"
inputs = tokenizer(input_text, return_tensors="pt")
outputs = model.generate(**inputs, max_length=512, do_sample=True, temperature=0.7)
workout = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(workout)

Input Parameters

Training Phases:

  • weight_loss: Fat loss focused routines
  • muscle_gain: Strength and muscle building
  • cardio_improve: Cardiovascular fitness enhancement
  • maintenance: General fitness maintenance

Motivation Types:

  • medical_recommendation: Health-prescribed exercise
  • self_improvement: Personal development goals
  • competition: Athletic/competitive training
  • rehabilitation: Recovery and therapy
  • wellbeing: General wellness and mood

Special Situations:

  • pregnancy: Prenatal fitness routines
  • post_partum: Postpartum recovery workouts
  • injury_recovery: Rehabilitation exercises
  • chronic_condition: Adapted for chronic health issues
  • elderly_population: Senior-friendly routines
  • physical_limitation: Modified for disabilities
  • none: Standard training

Training Details

Training Data

The model was trained on synthetic workout data generated from multiple comprehensive exercise databases:

  1. Primary Dataset: onurSakar/GYM-Exercise - Structured gym exercise data with instructions and categories
  2. Kaggle Dataset: niharika41298/gym-exercise-data - Comprehensive exercise descriptions with difficulty levels
  3. GitHub Dataset: yuhonas/free-exercise-db - Open-source exercise database with detailed instructions

Total Training Examples: 2000+ synthetic workout scenarios covering diverse user contexts and exercise combinations.

Training Procedure

Preprocessing

  1. Data Extraction: Parsed exercise data from multiple sources using regex patterns
  2. Data Normalization: Standardized column names and formats across datasets
  3. Synthetic Generation: Created workout scenarios using template-based generation
  4. Context Mapping: Paired user contexts with appropriate workout structures
  5. Tokenization: Applied FLAN-T5 tokenizer with padding and truncation

Training Hyperparameters

  • Base Model: google/flan-t5-base (850M parameters)
  • Training Epochs: 1 per phase
  • Batch Size: 8 (original) / 1 (optimized with gradient accumulation)
  • Gradient Accumulation Steps: 8
  • Learning Rate: 5e-5
  • Weight Decay: 0.01
  • Training Regime: FP16 mixed precision
  • Optimizer: AdamW
  • Evaluation Strategy: Steps-based (every 200/500 steps)
  • Save Strategy: Every 500/1000 steps

Training Architecture

Multi-Phase Training Approach:

  1. Phase 1: Fine-tuning on Kaggle-derived synthetic data (1000 examples)
  2. Phase 2: Additional training on GitHub-derived synthetic data (1000 examples)

Speeds, Sizes, Times

  • Model Size: ~850MB (FLAN-T5-base architecture)
  • Training Time: ~2-4 hours per phase on Google Colab GPU
  • Inference Speed: 0.5-1 seconds per workout (GPU), 2-5 seconds (CPU)
  • Memory Requirements: 4-8GB RAM for inference, 12GB+ for training

Evaluation

Testing Data, Factors & Metrics

Testing Data

The model was evaluated on held-out synthetic workout data representing diverse user contexts and exercise combinations. Test cases included edge cases such as multiple special situations and complex user requirements.

Factors

Evaluation was performed across multiple factors:

  • Training Phase Diversity: All four training phases (weight_loss, muscle_gain, cardio_improve, maintenance)
  • Motivation Variety: Five motivation types from medical to competitive
  • Special Situations: Seven different special health/physical situations
  • Workout Complexity: Varying block structures and exercise counts

Metrics

Primary evaluation metrics included:

  • JSON Format Validity: Structural correctness of generated workout plans
  • Context Relevance: Appropriateness of exercises for given user context
  • Exercise Variety: Diversity in recommended exercises across similar contexts
  • Block Structure Coherence: Logical flow from warmup to cooldown
  • Duration Estimation Accuracy: Realistic time estimates for workout completion

Results

The model successfully generates valid JSON workout structures with contextually appropriate exercises. Manual evaluation showed strong performance in:

  • Context understanding and exercise selection
  • Workout structure and flow
  • Adaptation to special situations
  • Exercise variety and avoiding repetition

Environmental Impact

Training AthenAI involved fine-tuning a pre-existing model rather than training from scratch, significantly reducing computational requirements and carbon footprint.

  • Hardware Type: NVIDIA T4 GPU (Google Colab)
  • Hours used: Approximately 6-8 hours total training time
  • Cloud Provider: Google Cloud Platform (Colab)
  • Compute Region: Variable (Colab allocation)
  • Carbon Emitted: Estimated <5kg CO2eq (due to fine-tuning approach)

Technical Specifications

Model Architecture and Objective

  • Architecture: FLAN-T5-base (Text-to-Text Transfer Transformer)
  • Parameters: ~850 million parameters
  • Objective: Sequence-to-sequence generation for workout plan creation
  • Input Format: Natural language context description
  • Output Format: Structured JSON workout plans
  • Context Window: 512 tokens maximum
  • Generation Strategy: Autoregressive text generation with temperature sampling

Compute Infrastructure

Hardware

  • Training: NVIDIA T4 GPU (Google Colab Pro)
  • Memory: 16GB GPU memory, 25GB system RAM
  • Storage: 100GB+ for datasets and model checkpoints

Software

  • Framework: Hugging Face Transformers 4.45.2
  • Training Library: Hugging Face Trainer
  • Data Processing: Pandas, Datasets 3.0.1
  • Environment: Python 3.10, PyTorch 2.0+
  • Platform: Google Colab with GPU acceleration

Citation

If you use AthenAI in your research or applications, please cite:

BibTeX:

@misc{athenai2024,
  title={AthenAI: AI-Powered Personalized Workout Generator},
  author={a-albiol},
  year={2024},
  publisher={Hugging Face},
  journal={Hugging Face Model Hub},
  howpublished={\url{https://huggingface.co/a-albiol/AthenAI}}
}

APA:

a-albiol. (2024). AthenAI: AI-Powered Personalized Workout Generator. Hugging Face Model Hub. https://huggingface.co/a-albiol/AthenAI

Glossary

  • Training Phase: User's current fitness goal (weight loss, muscle gain, etc.)
  • Motivation: Underlying reason for exercising (medical, competition, etc.)
  • Special Situation: Health or physical considerations (pregnancy, injury, etc.)
  • Workout Block: Structured section of workout (warmup, main, cooldown)
  • Fine-tuning: Process of adapting pre-trained model to specific task
  • Synthetic Data: Artificially generated training examples based on real exercise databases

More Information

For detailed implementation, training notebooks, and additional examples, visit the project repository. The model continues to evolve with community feedback and additional training data.

For technical support or questions about integration, please open an issue in the repository or contact through Hugging Face model discussions.

Model Card Authors

Primary Author: a-albiol

Contributors: Community feedback and testing

Model Card Contact

For questions, feedback, or collaboration inquiries:

  • Hugging Face: @a-albiol
  • Model Discussions: Use the Hugging Face model page discussion section
  • Issues: Report technical issues through the repository issue tracker
Downloads last month
110
Safetensors
Model size
248M params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for a-albiol/AthenAI

Unable to build the model tree, the base model loops to the model itself. Learn more.

Dataset used to train a-albiol/AthenAI