---
library_name: transformers
tags:
  - trl
  - sft
  - fine-tuned
  - custom-dataset
  - navigation-ai
  - coach-assistant
  - llama
license: apache-2.0
language:
  - en
base_model: meta-llama/Llama-3.2-3B-Instruct
---

# Coach Navigation Assistant — Fine-Tuned Model

This model is fine-tuned to help coaches navigate a dashboard by generating helpful paragraphs that include correct internal HTML links (e.g., `<a href="/courses/add">Click here</a>`). It is designed for use inside LMS/admin dashboards where users need quick instructions and direct links.

## Model Details

### Model Description

This model is a fine-tuned version of **meta-llama/Llama-3.2-3B-Instruct**. Its purpose is to understand natural language questions such as:

- "Where do I go to create a course?"
- "How can I add a new video section?"
- "Where do I manage students?"

And return a friendly paragraph that explains the location and includes the correct internal endpoint as a clickable HTML link.

- **Developed by:** Ziyad T.
- **Model type:** Instruction-tuned generative model
- **Languages:** English
- **Finetuned from:** meta-llama/Llama-3.2-3B-Instruct
- **License:** Apache 2.0 (same as base model)

### Model Sources

- **Repository:** [This Hugging Face model page]
- **Demo:** [Optional - Add your demo link]
- **Training Notebook:** [Add your Colab/Kaggle link]

## Uses

### Direct Use

The model can be used to:

- Generate dashboard navigation guidance
- Provide contextual instructions
- Embed HTML links inside responses
- Power chatbots for course creators, admins, or coaches

### Downstream Use

- LMS support assistants
- Platform onboarding bots
- Context-aware help centers
- Interactive documentation systems

### Out-of-Scope Use

The model is **NOT** suitable for:

- Factual Q&A outside the navigation domain
- Sensitive or medical advice
- Arbitrary text generation not related to navigation
- Producing links outside your controlled system
- General-purpose conversation

## Bias, Risks, and Limitations

⚠️ **Important Limitations:**

- The model will only work correctly with endpoints it was trained on
- If the user asks about a route not in the dataset, the model may guess or hallucinate a link
- The model assumes the platform uses HTML `<a>` tags — not markdown or other formats
- Responses are optimized for coach/admin users, not students or general users

### Recommendations

- ✅ Validate model outputs before exposing them publicly
- ✅ Keep endpoints consistent with the dataset
- ✅ Add more training examples as your app grows
- ✅ Implement fallback responses for unknown queries
- ✅ Monitor and log model outputs for quality assurance

## How to Use the Model

### Basic Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("your-username/llama-3.2-3b-coach-assistant")
model = AutoModelForCausalLM.from_pretrained("your-username/llama-3.2-3b-coach-assistant")

# Prepare input
prompt = "Where do I go to create a new course?"

# Generate response
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=200, temperature=0.7)

# Decode output
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)
```

### Using with Hugging Face Inference API

```python
import requests

API_URL = "https://api-inference.huggingface.co/models/your-username/llama-3.2-3b-coach-assistant"
headers = {"Authorization": "Bearer YOUR_HF_TOKEN"}

def query(payload):
    response = requests.post(API_URL, headers=headers, json=payload)
    return response.json()

output = query({
    "inputs": "Where do I go to create a new course?",
    "parameters": {
        "max_new_tokens": 150,
        "temperature": 0.7,
    }
})

print(output)
```

### Using with Inference Endpoints (Recommended for Production)

```python
import requests

# Your deployed endpoint URL
API_URL = "https://your-endpoint.aws.endpoints.huggingface.cloud"
headers = {"Authorization": "Bearer YOUR_HF_TOKEN"}

payload = {
    "inputs": "Where do I go to create a new course?",
    "parameters": {
        "max_new_tokens": 150,
        "temperature": 0.7,
        "top_p": 0.9,
    }
}

response = requests.post(API_URL, headers=headers, json=payload)
print(response.json())
```

## Training Details

### Training Data

A custom JSONL dataset consisting of input-output pairs:

```json
{
  "input": "Where do I go to create a new course?",
  "output": "To create a new course, navigate to the course creation page where you can add all the details about your course. <a href=\"/courses/add\">Click here</a> to start creating your course."
}
```

**Dataset Statistics:**
- Total examples: 50+
- Endpoints covered: 8+ major routes
- Variations per endpoint: 3-5 different question phrasings

**Covered Endpoints:**
- `/courses/add` - Creating new courses
- `/courses/index` - Viewing and managing courses
- `/inbox` - Checking messages
- `/accont` - Managing profile
- `/coach/dashboard` - Accessing dashboard
- Course editing, sections, students, and more

### Training Procedure

**Fine-tuning Method:** QLoRA (4-bit quantization + LoRA)

**Training Framework:**
- TRL SFTTrainer (Supervised Fine-Tuning)
- PEFT (Parameter-Efficient Fine-Tuning)
- BitsAndBytes (4-bit quantization)

**Training Configuration:**
- **Base Model:** meta-llama/Llama-3.2-3B-Instruct
- **Quantization:** 4-bit NF4 with double quantization
- **LoRA Rank (r):** 16
- **LoRA Alpha:** 32
- **LoRA Dropout:** 0.05
- **Target Modules:** q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
- **Learning Rate:** 2e-4
- **Batch Size:** 4 per device
- **Gradient Accumulation Steps:** 4 (effective batch size: 16)
- **Epochs:** 3
- **Max Sequence Length:** 512
- **Optimizer:** paged_adamw_8bit
- **LR Scheduler:** Cosine
- **Warmup Steps:** 50
- **Mixed Precision:** FP16

### Speeds, Sizes, Times

**Model Size:**
- Base model parameters: ~3B
- Trainable parameters (LoRA): ~0.5% of total
- Final model size: ~6GB (merged)

**Training Environment:**
- GPU: [Add your GPU type, e.g., NVIDIA A100, T4]
- Training time: [Add your training time, e.g., ~2 hours]
- Cloud platform: [Add if applicable, e.g., Google Colab, Kaggle, AWS]

## Evaluation

### Testing Data

A validation set with unseen question variations for each endpoint to test generalization.

### Metrics

**Evaluation Criteria:**
- ✅ **Link Correctness:** Does the model return the correct endpoint?
- ✅ **Response Quality:** Is the paragraph helpful and natural?
- ✅ **Format Consistency:** Does it follow the HTML link format?
- ✅ **Instruction Clarity:** Are the instructions clear and actionable?

### Results

The model reliably returns:
- ✅ The correct link for trained endpoints
- ✅ Relevant and contextual instructions
- ✅ Consistent HTML formatting
- ✅ Natural, conversational language

**Sample Outputs:**

| Input | Output |
|-------|--------|
| "Where do I create a course?" | "To create a new course, navigate to the course creation page where you can add all the details about your course. <a href=\"/courses/add\">Click here</a> to start creating your course." |
| "How can I check my messages?" | "You can check all your messages and communicate with students in the inbox. <a href=\"/inbox\">Click here</a> to access your messages." |
| "Where is my dashboard?" | "Your coach dashboard provides an overview of your courses, students, and activity. <a href=\"/coach/dashboard\">Click here</a> to access your dashboard." |

## Environmental Impact

**Carbon Emissions:** [Optional - Add if you tracked this]

Fine-tuning was performed using cloud GPU resources. The use of QLoRA (4-bit quantization) significantly reduced computational requirements compared to full fine-tuning.

## Technical Specifications

### Model Architecture

- **Architecture:** Llama 3.2 (decoder-only transformer)
- **Parameters:** ~3 billion
- **Attention:** Multi-head attention with grouped-query attention
- **Context Length:** 8192 tokens (base model capability)
- **Training Context:** 512 tokens (for efficiency)

### Compute Infrastructure

- **Hardware:** [Add your GPU type]
- **Software:** 
  - Python 3.10+
  - PyTorch 2.0+
  - Transformers 4.35+
  - PEFT 0.6+
  - BitsAndBytes 0.41+
  - TRL (Transformer Reinforcement Learning)

## Citation

If you use this model in your work, please cite:

**BibTeX:**
```bibtex
@misc{coach-assistant-2024,
  author = {Ziyad T.},
  title = {Coach Navigation Assistant - Fine-Tuned Llama 3.2 3B},
  year = {2024},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/your-username/llama-3.2-3b-coach-assistant}}
}
```

**APA:**
```
Ziyad T. (2024). Coach Navigation Assistant - Fine-Tuned Llama 3.2 3B. 
Hugging Face. https://huggingface.co/your-username/llama-3.2-3b-coach-assistant
```

## Model Card Authors

**Ziyad T.**

## Model Card Contact

For questions, issues, or feedback:
- **GitHub:** [Add your GitHub profile]
- **Email:** [Add your email]
- **Hugging Face:** [Add your HF profile]

## Acknowledgments

- Base model: Meta AI (Llama 3.2)
- Training framework: Hugging Face (Transformers, PEFT, TRL)
- Quantization: BitsAndBytes team

---

**Last Updated:** November 2024