Configuration structure

Eole validates configuration through Pydantic models. Unknown fields are rejected. Use a YAML file with eole train -c train.yaml or eole predict -c predict.yaml. The command runner expands environment variables in YAML paths.

A training file has top-level corpus/vocabulary/logging settings and nested model, training, and transforms_configs sections. For fine-tuning, training.train_from supplies the saved architecture and tokenization settings. A prediction file places model_path, src, device, precision, and decoding settings at the top level; compatible settings are loaded from the checkpoint.

model_path: /path/to/converted/model
src: input.txt
output: output.txt
gpu_ranks: [0]
compute_dtype: bf16
beam_size: 1
top_k: 1

Top-level run fields can be overridden by their CLI flags. Nested fields should be edited in YAML; dotted CLI overrides such as --training.compute_dtype are not supported. Server YAML (models:) and recipe orchestration YAML are separate schemas, not prediction configurations.

See the configuration pages below for individual fields, and the website recipe index for complete model/data examples. GPU kernel availability is independent of schema validation.