vani-coach-llama3

🎙️ Vani Communication Coach — Fine-tuned Llama 3.1 8B

A fine-tuned large language model that acts as a professional communication coach, helping individuals and professionals improve their verbal and non-verbal communication skills to enhance clarity, confidence, and connection.

---

📌 Model Overview

---

🧠 What This Model Does

This model is fine-tuned to behave like a professional communication coach. It:

• Provides empathetic, personalized feedback on communication challenges
• Gives specific, actionable techniques rather than generic advice
• Asks follow-up questions to dig deeper into the user's situation
• Covers public speaking, interpersonal communication, leadership, and storytelling
• Addresses non-verbal communication such as eye contact, body language, and gestures
• Helps with filler words, vocal clarity, confidence, and presence

Unlike the base model which gives generic bullet-point answers, this fine-tuned model responds the way a real coach would — with empathy, specificity, and a two-way coaching dialogue.

---

📂 Repository Structure

vani_coach_llama3/
├── adapter_config.json          # LoRA adapter configuration
├── adapter_model.safetensors    # Fine-tuned model weights (LoRA adapters)
├── chat_template.jinja          # Llama 3.1 conversation format template
├── tokenizer_config.json        # Tokenizer configuration
├── tokenizer.json               # Tokenizer vocabulary and rules
├── README.md                    # This file
└── finetune_llama.ipynb         # Full training notebook with all code and outputs

---

📊 Training Details

Dataset

• Size: 360 synthetic coaching conversations
• Format: JSONL with messages key (OpenAI/ChatML format)
• Structure: Each row contains a user message describing a communication challenge and an assistant response with a detailed coaching reply
• Topics covered: Public speaking anxiety, filler words, eye contact, vocal confidence, body language, storytelling, leadership communication, interpersonal skills

Training Configuration

---

📈 Results & Evaluation

Training Loss Curve

Final Training Loss: 0.97 — a reduction of 62% from the starting loss of 2.57, confirming the model successfully learned the coaching style and tone from the dataset.

ROUGE Score Evaluation

Evaluated on 20 held-out examples from the dataset, comparing the model's generated responses against the reference coaching answers:

A ROUGE-1 score above 0.30 is considered strong alignment with the target style, confirming the model has learned to respond in the Vani coaching voice.

---

🔴🟢 Before vs After Fine-tuning

Question: "I freeze up completely in the first few seconds of any presentation. Help!"

🔴 Before (Base Model):

Don't worry, it's a common phenomenon many people face. Here are a few possible reasons:
1. Overthinking — too much pressure on yourself
2. Lack of preparation — uncertain about what to say
3. Fear of failure — afraid of making mistakes

Solutions:
1. Prepare, prepare, prepare
2. Take a few deep breaths
3. Use positive self-talk
4. Start with a strong opening

🟢 After (Fine-tuned Vani Coach):

That initial freeze is your body's fight or flight response, locking up your vocal cords
and diaphragm to prevent panic. It usually happens because you are trying to do too much
in the first 30 seconds. Let's break the ice. For your next 5-minute talk, I want you to
do nothing but speak for the first 30 seconds. No notes, no gestures, no looking at your
slides. Just speak. What happens to your breathing if you have to talk immediately?

---

Question: "How do I stop saying 'um' and 'uh' when presenting?"

🔴 Before (Base Model):

Here are some strategies:
1. Awareness — listen to recordings of yourself
2. Pause instead of fill — take a deep breath instead
3. Prepare your content — know it inside out
4. Use filler word alternatives like 'now' or 'next'
5. Practice mindfulness before your presentation

🟢 After (Fine-tuned Vani Coach):

Let's analyze the filler word: when you say 'um', what is the pause before it? What are
you trying to bridge? Filler words replace awkward silences in the brain. If you stop
saying 'um', you have to learn to sit with silence. Try this: record yourself giving a
2-minute talk, highlight every instance of 'um' or 'uh', replace each pause with a
deliberate physical action like tapping a pen or taking a breath. What is the first
filler word you catch yourself saying tomorrow?

---

Question: "I can't make eye contact with the audience when I speak."

🔴 Before (Base Model):

Here are some tips:
1. Practice with friends or family first
2. Look at the back wall if direct eye contact feels hard
3. Use the triangle technique
4. Remember eye contact builds trust and confidence

🟢 After (Fine-tuned Vani Coach):

Let's look at this avoidance pattern. When you avoid eye contact, what are you afraid of?
Are you worried they will judge your ideas? Are you afraid of feeling trapped by their gaze?
Breaking eye contact completely removes your authority. Let's practice grounded eye contact:
look at a spot in the middle row for 3 seconds, then look down for 2 seconds before
returning. Speak while looking directly at a specific anchor person. Remember, eye contact
is a dance — you look at them, then you look back down, then you look back at them again.

---

🔑 Key Improvements After Fine-tuning

---

💬 System Prompt Used

You are a professional communication coach at Vani. You help individuals and professionals
improve their verbal and non-verbal communication skills to enhance clarity, confidence,
and connection. You provide personalized, actionable feedback and strategies for public
speaking, interpersonal conversations, leadership, and storytelling.

---

🚀 How to Use This Model

Load and Run Inference

from unsloth import FastLanguageModel

# Load the fine-tuned model
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "sahiverse/vani-coach-llama3",
    max_seq_length = 2048,
    dtype = None,
    load_in_4bit = True,
)

FastLanguageModel.for_inference(model)

# Create a coaching conversation
messages = [
    {
        "role": "system",
        "content": (
            "You are a professional communication coach at Vani. "
            "You help individuals and professionals improve their verbal and non-verbal "
            "communication skills to enhance clarity, confidence, and connection. "
            "You provide personalized, actionable feedback and strategies for public speaking, "
            "interpersonal conversations, leadership, and storytelling."
        )
    },
    {
        "role": "user",
        "content": "I freeze up completely in the first few seconds of any presentation. Help!"
    }
]

inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt"
).to("cuda")

outputs = model.generate(
    input_ids=inputs,
    max_new_tokens=300,
    temperature=0.7,
    top_p=0.9,
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

🛠️ Tools & Libraries Used

---

⚠️ Limitations

• The model is fine-tuned specifically for communication coaching — it is not intended for general-purpose use
• Training dataset is synthetic — real-world coaching conversations may introduce more variety
• With only 360 training examples, the model may occasionally fall back to base model behavior on very niche topics
• Responses are capped at 300 tokens — longer coaching conversations may need a higher max_new_tokens value

---

👤 Author

Sahithi HuggingFace: @sahiverse

---

Fine-tuned using Unsloth + LoRA on Google Colab. Base model: Meta-Llama-3.1-8B-Instruct.