How furkankirat/konjin-ai works
konjin-ai is a small PyTorch byte-level GPT learning project with two product workflows: generate C# boilerplate and conduct Turkish chat. Data preparation feeds model training; trained checkpoints are used by interactive applications and boilerplate evaluation. The model, training loop, data pipeline and sampling live in one file, so those responsibilities share its exact path. Edges reflect the README and shown imports/calls; unsampled implementation details are not inferred.
Main entry points: minigpt.py Training and sampling · bp_demo.py Boilerplate demo · gen_boilerplate.py C# example generator · prep_tr.py Turkish data preparation
Components 11
Data preparation
- C# example generator
src/datagen/gen_boilerplate.py - Turkish data preparation
src/datagen/prep_tr.py
Model and training
- Training and sampling
src/model/minigpt.py - Byte-level GPT
src/model/minigpt.py - Model checkpoint
User applications
- Boilerplate demo
src/eval/bp_demo.py - Turkish chat REPL
src/model/tr_chat.py
Boilerplate evaluation
- Held-out evaluator
src/eval/bp_eval.py - Exact-match scorer
src/eval/bp_check.py
Other components
- User
- Training datasets
Connections 16
- User → C# example generator: requests data
README.md - C# example generator → Training and sampling: supplies examples
README.md - User → Turkish data preparation: prepares corpus
README.md - Training datasets → Turkish data preparation: provides source data
src/datagen/prep_tr.py - Turkish data preparation → Training and sampling: writes training text
README.md - Training and sampling → Byte-level GPT: builds and trains
src/model/minigpt.py - Training and sampling → Model checkpoint: saves weights
README.md - User → Boilerplate demo: provides class fields
src/eval/bp_demo.py - Boilerplate demo → C# example generator: builds reference output
src/eval/bp_demo.py - Boilerplate demo → Training and sampling: loads and samples
src/eval/bp_demo.py - Model checkpoint → Boilerplate demo: loads model
README.md - User → Turkish chat REPL: sends messages
src/model/tr_chat.py - Turkish chat REPL → Training and sampling: loads and samples
src/model/tr_chat.py - Model checkpoint → Turkish chat REPL: loads model
README.md - Held-out evaluator → Training and sampling: loads and samples
src/eval/bp_eval.py - Exact-match scorer → User: reports exact match
src/eval/bp_check.py