How celiktepesahin/chatgpt-from-scratch-tr works
This repository is a Turkish-language course, delivered as one notebook, that guides a learner from tokenizer and tensor concepts through GPT-2 construction, training, and fine-tuning. The README documents the major stages and data sources, but no notebook source was sampled, so the graph reflects documented course flow and capabilities rather than verified internal calls.
Main entry points: FULL-SERIES.ipynb Weight handling · instruction-data.json Instruction dataset · how-words-work.txt Pretraining corpus
Components 16
Course and data
- Course notebook
FULL-SERIES.ipynb - Pretraining corpus
_data/how-words-work.txt - Instruction dataset
_data/instruction-data.json
Model construction
- BPE tokenizer
FULL-SERIES.ipynb - Tensor pipeline
FULL-SERIES.ipynb - Attention
FULL-SERIES.ipynb - Causal multi-head
FULL-SERIES.ipynb - Network layers
FULL-SERIES.ipynb - GPT-2 blocks
FULL-SERIES.ipynb
Training and generation
- Pretraining loop
FULL-SERIES.ipynb - Weight handling
FULL-SERIES.ipynb - Text generation
FULL-SERIES.ipynb
Fine-tuning
- Topic classifier
FULL-SERIES.ipynb - Instruction tuning
FULL-SERIES.ipynb - Assistant evaluation
FULL-SERIES.ipynb
Other components
- Learner
Connections 15
- Learner → Course notebook: runs lessons
README.md - Course notebook → Pretraining corpus: searches for data
README.md - Course notebook → Instruction dataset: uses for tuning
README.md - BPE tokenizer → Tensor pipeline: prepares tokens
README.md - Tensor pipeline → Attention: feeds examples
- Attention → Causal multi-head: extends to
README.md - Causal multi-head → Network layers: combines with
- Network layers → GPT-2 blocks: forms blocks
README.md - GPT-2 blocks → Pretraining loop: trains model
README.md - Pretraining loop → Weight handling: saves weights
README.md - Weight handling → Text generation: supports sampling
README.md - Weight handling → Topic classifier: fine-tunes for classification
README.md - Instruction dataset → Instruction tuning: supplies examples
README.md - Weight handling → Instruction tuning: fine-tunes for instructions
README.md - Instruction tuning → Assistant evaluation: evaluates answers
README.md