Configure Model Endpoints for BYOB

BYOB does not ship model weights. Instead, YAML blocks describe OpenAI-compatible clients for each stage that calls a language model.

Generation and judgement

generation_model_config and judge_model_config are required mappings. They follow the same structure you see in src/nemotron/steps/byob/mcq/config/default.yaml: alias, model, provider, and nested inference_parameters such as max_tokens, max_parallel_requests, temperature, and top_p.

setup_model_config in runtime/data_designer_utils.py reads these blocks when Data Designer runs batched stages.

Distractor and validity stages

When do_distractor_expansion is true, distractor_expansion_model_config must be present. distractor_validity_model_config is always required in the current schema validation.

Filtering swarm

filtering_model_configs contains two lists, hallucination and easiness. Each list entry needs a unique alias across both lists. The structure mirrors other model blocks so you can aim filters at different endpoints than the generator.

Translation models

Translation uses translation_model_config in translate.yaml. backend_type, params (including api_key_env and base_url), and nested stage / segment_stage keys are passed into Curator’s experimental translation pipeline.

API keys

Model blocks may set api_key_env to read secrets from the environment (for example NGC_API_KEY in the sample translation config). Set those variables in the shell that launches nemotron steps run.