hmtr/rolebox.coach-kettlebell

RoleBox Fitness Advisor (Coach Kettlebell)

A specialized fitness advisor AI trained to provide personalized fitness guidance, workout recommendations, and exercise advice.

Model Details

Model Description

This is a LoRA adapter fine-tuned on top of Qwen 2.5 Coder 1.5B Instruct to create a specialized fitness advisor. The model provides evidence-based fitness advice, workout recommendations, exercise form guidance, and answers fitness-related questions.

• Developed by: RoleBox Team
• Model type: Causal Language Model (LoRA adapter)
• Language(s): English
• License: Apache 2.0
• Finetuned from model: Qwen/Qwen2.5-Coder-1.5B-Instruct

Model Sources

• Parent Model: https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B-Instruct

Uses

Direct Use

This model is designed to be used as a fitness advisor chatbot. It can:

• Provide workout recommendations based on goals (muscle building, weight loss, endurance)
• Explain proper exercise form and technique
• Suggest training programs and routines
• Answer questions about fitness, nutrition, and recovery
• Offer motivation and guidance for fitness journeys

Example Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-Coder-1.5B-Instruct",
    torch_dtype=torch.float16,
    device_map="auto"
)

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-Coder-1.5B-Instruct")

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "hmtr/rolebox.coach-kettlebell")

# Generate response
prompt = """### Instruction:
You are a fitness advisor. Answer the user's question.

### User Question:
I want to build muscle. What exercises should I do?

### Response:
"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.7)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Downstream Use

This adapter can be integrated into:
- Fitness coaching applications
- Health and wellness platforms
- Workout planning tools
- Conversational AI assistants for gyms and fitness centers

Out-of-Scope Use

This model should NOT be used for:
- Medical diagnosis or treatment recommendations
- Personalized medical advice (consult healthcare professionals)
- Nutritional prescriptions for medical conditions
- Rehabilitation from injuries (requires professional guidance)
- Replacing professional personal trainers for complex needs

Important: Always consult with healthcare professionals before starting any new fitness program, especially if you have pre-existing conditions.

Bias, Risks, and Limitations

- Training data bias: The model was trained on general fitness Q&A data and may not represent all fitness philosophies or training methods
- Not personalized: Cannot account for individual health conditions, injuries, or limitations
- General advice only: Recommendations are general and not tailored to individual fitness levels
- No medical training: The model is not a substitute for professional medical or fitness advice
- English only: Currently trained only on English-language fitness content

Recommendations

- Users should verify fitness advice with certified personal trainers
- Always start with lower intensity and progress gradually
- Consult healthcare providers before beginning new exercise programs
- Use the model as a supplement to, not a replacement for, professional guidance
- Be cautious with form-critical exercises (consult trainers for proper technique)

Training Details

Training Data

The model was fine-tuned on a curated dataset of 965 fitness-related question-answer pairs covering:
- Strength training and muscle building
- Cardiovascular exercise
- Flexibility and mobility
- Exercise form and technique
- Workout programming
- Recovery and rest
- Common fitness questions

Training Procedure

Fine-tuning method: LoRA (Low-Rank Adaptation)

Training Hyperparameters

- Base model: Qwen/Qwen2.5-Coder-1.5B-Instruct
- Training regime: fp16 mixed precision
- LoRA rank (r): 16
- LoRA alpha: 32
- LoRA dropout: 0.05
- Target modules: q_proj, k_proj, v_proj, o_proj
- Number of epochs: 3
- Batch size: 4
- Gradient accumulation steps: 2 (effective batch size: 8)
- Learning rate: 2e-4
- Max sequence length: 384 tokens
- Optimizer: AdamW
- Training examples: 965

Speeds, Sizes, Times

- Adapter size: ~52 MB
- Training time: ~10-15 minutes on Google Colab T4 GPU
- Training platform: Google Colab (free tier)
- GPU: NVIDIA Tesla T4 (16GB VRAM)
- Trainable parameters: ~13.6M (0.9% of base model)

Evaluation

This model was evaluated through qualitative testing on fitness-related queries. Quantitative metrics to be added in future versions.

Testing Data

Held-out fitness Q&A examples covering diverse topics:
- Workout routines
- Exercise technique
- Training principles
- Recovery strategies

Metrics

- Qualitative assessment of response relevance and accuracy
- Manual review by fitness domain experts (planned)

Environmental Impact

Training was performed on Google Colab's free tier GPU infrastructure.

- Hardware Type: NVIDIA Tesla T4 GPU
- Hours used: ~0.2 hours (10-15 minutes)
- Cloud Provider: Google Cloud Platform
- Compute Region: US (variable)
- Carbon Emitted: Minimal (<0.01 kg CO2eq estimated)

Technical Specifications

Model Architecture and Objective

- Architecture: Transformer-based causal language model with LoRA adapters
- Objective: Causal language modeling (next token prediction)
- Adapter method: LoRA (Low-Rank Adaptation)
- Parameter efficiency: Only 0.9% of parameters are trainable

Compute Infrastructure

Hardware

- Training: Google Colab T4 GPU (16GB VRAM)
- Inference: Can run on consumer GPUs (4GB+ VRAM) or CPU

Software

- Framework: PyTorch
- Libraries:
  - Transformers (Hugging Face)
  - PEFT (Parameter-Efficient Fine-Tuning)
  - Accelerate
  - Datasets

Citation

BibTeX:

@misc{rolebox-fitness-advisor,
  title={RoleBox Fitness Advisor: LoRA-finetuned Qwen 2.5 Coder for Fitness Guidance},
  author={RoleBox Team},
  year={2025},
  publisher={HuggingFace},
  url={https://huggingface.co/hmtr/rolebox.coach-kettlebell}
}

Model Card Authors

RoleBox Team

Model Card Contact

- Email: [email protected]
- Website: https://rolebox.app

Framework Versions

- PEFT 0.17.1
- Transformers 4.48+
- PyTorch 2.6+
- Python 3.10+