Orinth
GitHubDownload

Models & Training

Orinth runs long training jobs outside the API request path. Jobs execute in subprocesses, which keeps the app responsive and the run observable.

The Model Library

The Model Library is the catalog of runnable bases:

  • Computer vision — Keras Applications (MobileNetV2, EfficientNet B0–B7, ResNet50, and others) and Ultralytics YOLO families (YOLO11, YOLO26).
  • NLP — Hugging Face transformers (BERT, BART) and Keras text classifiers.
  • LLM — Hugging Face checkpoints from the Hub or a local upload.

It also holds what you produce: trained models promoted out of runs, and weights you upload yourself.

Training workflow

  • Task first — training begins by selecting a task (object detection, summarization, and so on). Model options and hyperparameters filter down to the choices compatible with your dataset.
  • Observable jobs — a running job streams logs in real time, parsing metrics into epoch-by-epoch progress and an estimated completion time.
  • Model promotion — completed YOLO and Keras runs are registered as stable artifacts automatically. Promoted models are immediately selectable for Testing and Inference.

Advanced settings and device detection

  • Advanced parameters — beyond epochs, batch size, and learning rate, an advanced accordion exposes each family's catalog defaults. Every value is submitted with the run, so a job is reproducible from its record alone.
  • Guardrails — the learning-rate field warns when a value drifts outside the sane band for the selected base. Fine-tuning a transformer lives in a much narrower band than training from scratch.
  • Auto-detected accelerator — vision and LLM runs read the available device (CUDA, Apple MPS, or CPU) and default to it, so the configuration you see is the one that will run.

LLM fine-tuning

LLM fine-tuning runs alongside the vision and NLP families. Choose an llm_finetune task, a Hugging Face base, and a method — LoRA, QLoRA, full fine-tune, or continued pretraining. See LLM Fine-tuning & Chat for the full walkthrough, including serving and GGUF export.

Artifacts

All logs, generated weights (best.pt, best_model.keras), configuration files, and results.csv files are written under local ignored directories. Manual promotion is idempotent, for the edge cases where automatic promotion does not fire.