🧠 Custom Knowledge LLM: Tony Stark Edition

This is a fine-tuned version of the Qwen2.5-3B-Instruct model, adapted to answer domain-specific questions related to Tony Stark, using the LoRA (Low-Rank Adaptation) method for parameter-efficient fine-tuning.

📌 Model Details

Model Description

This project is a fun + educational experiment that fine-tunes a base LLM using a fictional dataset based on Tony Stark from the Marvel universe.

Developed by: Aviral Srivastava
Model type: Causal Language Model (Instruction-tuned)
Language: English
License: MIT
Finetuned from model: Qwen/Qwen2.5-3B-Instruct

🧑‍💻 Uses

Direct Use

This model is fine-tuned to answer Tony Stark–related prompts such as:

"Who is Tony Stark?"
"What suits did Iron Man build?"
"What are leadership traits of Stark?"

Downstream Use

The methodology can be directly reused for:

Corporate knowledge assistants
Domain-specific customer support
Educational tutors trained on custom material
Healthcare, law, and e-commerce Q&A bots

Out-of-Scope Use

This model is not designed for:

Real-world advice in medical, legal, or financial domains
Factual accuracy outside of Tony Stark lore
Handling unrelated general-purpose queries

⚠️ Bias, Risks, and Limitations

This model is trained on fictional data and is not meant for serious tasks.
It reflects only the content provided in the custom dataset.
It may "hallucinate" facts if asked general questions.

Recommendations

Please do not use this for any commercial or factual purpose without re-training on a verified dataset.

🚀 How to Use

from transformers import pipeline

qa = pipeline(
    model="Avirallm/Custom-Knowledge-LLM-Tony-Stark-Edition",
    tokenizer="Avirallm/Custom-Knowledge-LLM-Tony-Stark-Edition",
    device="cuda"  # or "cpu" if not using GPU
)

qa("List all Iron Man suits and their features.")

🏋️‍♂️ Training Details

📦 Training Data

A custom JSON dataset of prompt-completion pairs related to Tony Stark. Example entry:

~json { "prompt": "Who is Tony Stark?", "completion": "Tony Stark is a fictional billionaire inventor from Marvel..." } ~

🔧 Training Hyperparameters

Epochs: 10
Batch Size: 1
Optimizer: AdamW
Learning Rate: 0.001
Mixed Precision: FP16
Framework: Hugging Face Trainer + PEFT LoRA

🖥️ Training Setup

Trained fully on Google Colab Free Tier
Using Qwen/Qwen2.5-3B-Instruct with LoRA adapters
Fine-tuned only adapter layers (not full model)

📊 Evaluation

This project is primarily exploratory and not evaluated on public benchmarks.

🌱 Environmental Impact

Hardware: Google Colab Free GPU (Tesla T4)
Training Time: ~380 seconds (10 epochs, 1580 steps)
Carbon Emission: Negligible (low-compute, single GPU)

🧠 Architecture

Base Model: Qwen2.5-3B-Instruct (Alibaba Cloud)
Fine-Tuning: LoRA adapters on top of base weights
Task Type: Text generation, instruction following
Token Limit: 128 tokens (during training)

✨ Example Applications

Fan-based AI chatbot (Iron Man Assistant)
Fictional universe assistants for games and comics
Domain-specific tutors for educational platforms
Startup knowledge bots (replace "Tony Stark" with your brand)

📁 Repository Structure

adapter_model.safetensors – LoRA adapter weights
tokenizer_config.json, tokenizer.json, vocab.json – Tokenizer files
README.md – Project overview
training_args.bin – Training arguments
tonyst.json (optional) – Custom dataset (if shared)

📬 Get in Touch

Have a use case in mind? Want your own custom-trained LLM?
📧 Email: sriaviralnarain@gmail.com
🔗 LinkedIn: Aviral Srivastava
💻 GitHub: aviral-sri

🙏 Credits

Base Model: Qwen2.5-3B-Instruct
Fine-Tuning: PEFT + LoRA
Tools Used:
- Hugging Face Transformers
- Hugging Face Datasets
- Google Colab
- W&B for tracking

Inspired by: Marvel's Tony Stark (for learning only, non-commercial)

🪪 License

This project is licensed under the MIT License.
Feel free to modify, share, and build upon it.

Avirallm
/

Custom-Knowledge-LLM-Tony-Stark-Edition