PopularMajor Project
Artificial Intelligence
Face Recognition Attendance System
Camera-based attendance with liveness detection and analytics
- Python
- OpenCV
- FaceNet
- Flask
- +2
₹4,900₹5,999
4.8/301 reviews
Artificial Intelligence · Capstone
Build a Language Model from Scratch
Tokeniser, transformer blocks and training loop written in PyTorch
5-6 weeks build
Overview
Every part of a GPT-style model built by hand. You train a byte-pair encoding tokeniser, write causal self-attention, feed-forward layers and positional embeddings, then train the model on a public-domain text corpus with a proper training loop. Loss curves, perplexity and sampled text at each checkpoint show the model learning.
What makes it stand up
- 01Byte-pair encoding tokeniser implemented from scratch
- 02Causal multi-head self-attention written without library layers
- 03Training loop with warm-up, cosine decay and gradient clipping
- 04Perplexity tracking and sampled text at each checkpoint
- 05Temperature and top-k sampling for generation
Module breakdown
- 01
Tokeniser
BPE merge learning, encoding and decoding.
- 02
Attention
Scaled dot-product attention with a causal mask.
- 03
Model
Stacked transformer blocks with layer norm and residuals.
- 04
Training
Batching, optimiser schedule and checkpointing.
- 05
Generation
Sampling strategies and an interactive prompt script.
After this, you will be able to
- Derive scaled dot-product attention on the whiteboard
- Explain every line of your training loop
- Demo the model generating text from your own checkpoint
Technology stack
- Python
- PyTorch
- NumPy
- Matplotlib
- Level
- Capstone
- Domain
- Artificial Intelligence
- Build time
- 5-6 weeks
- Difficulty
You receive
- Complete, commented source code
- 60-90 page project report (IEEE format)
- Editable presentation deck
- Architecture, ER and UML diagrams
- Local setup and deployment guide
- Sample dataset or seed data
- Viva question bank with answers
- One 45-minute walkthrough call
- 30 days of doubt-clearing support