Configs

This directory contains YAML configs used for training, evaluation, and inference with Oumi, in addition to launching remote jobs for those tasks. We use Omegaconf to load these config files into our config classes.

As a convention, we name these files according to their corresponding class. For example:

• *train.yaml: TrainingConfig
• *eval.yaml: EvaluationConfig
• *infer.yaml: InferenceConfig
• *gcp_job.yaml: JobConfig, GCP cloud
• *polaris_job.yaml: JobConfig, Polaris cloud

Structure

We use the following sub-directories to organize our configs:

• recipes/: This directory contains configs for training, evaluation, and inference of common model families. This is a great starting point for most users.
• projects/: Configs for fully replicating the training of specific models (ex. Aya).
• examples/: Configs for specific use cases. This is less structured than configs/recipes and configs/projects, and includes more experimental jobs.