Training engine
Qwen 3.6 / Unsloth = domain adaptation, SFT, and GRPO. Full = nanoGPT (PyTorch). Tiny = microGPT (pure Python).
Data Sources
Connect S3, PostgreSQL, DWH (BigQuery, Redshift, Snowflake), or use local files. Credentials are stored securely.
Training runs
Data is trained in chunks to avoid overfitting and memory issues. Each run has a queue of chunks.
Training dashboards
Metrics and parameters per run. Status, chunks progress, and loss.
Chunk queue
Chunks are processed one by one (or in small batches). Current chunk and progress are shown below.
Run Full training
Starts nanoGPT train.py. Select your run to train on your data; otherwise uses Shakespeare demo.
Training parameters
If nanoGPT is missing: clone next to the app (git clone https://github.com/karpathy/nanoGPT ../nanoGPT) or set NANOGPT_PATH to the nanoGPT folder.
Samples (generate from trained nanoGPT)
Generate text from a checkpoint. Run Full training first or select an existing checkpoint.
Autotrain
Autonomous LLM training: the agent experiments with architecture, hyperparameters, and optimizers. Each iteration trains for a fixed time budget, evaluates val_bpb, keeps or discards changes.