Gym: Skill for Codex

Add a Model Config to Gym

The Gym pages are the reference; when they and this skill disagree, the pages win: Configuration (the services and driver fields), Multi-node vLLM and vLLM model server (the vllm_model fields, chat_template_kwargs, extra_body).

A model in a Gym evaluation has two parts:

• The deployment: a vllm service in the submit config, started with vllm serve, and named as the policy with driver.policy_model.
• The client: the model-type config Gym talks to it with (driver.policy_model_type, the --model-type of gym eval run). openai_model is the default; vllm_model adds reasoning extraction, chat_template_kwargs, extra_body and token information.

1. Read the model card first

Collect before writing anything: the context length, the recommended sampling settings, whether the model reasons (and which vLLM reasoning parser its family uses), its tool-call format (and the vLLM tool-call parser), whether its chat template has a reasoning switch and what it is called, and whether its repository ships a generation_config.json.

Without a model card, read the checkpoint itself: chat_template.jinja (or chat_template in tokenizer_config.json) for the reasoning switch and the tool-call format, generation_config.json for sampling, and config.json or tokenizer_config.json for the context length. The parser names come from the serving image: vllm serve --help=all lists the choices of --tool-call-parser and --reasoning-parser. Pick the tool-call parser whose format the template emits, and say in your notes which parser you chose and why.

If the benchmarks use tools, also check that the served template renders them. It needs a tools branch (grep -c tools chat_template.jinja is above 0) that renders the request's tools, the assistant's tool calls and the tool results. A template without one drops the tools without any error: the model answers in free text, the agent sees no tool call, and every rollout ends in a format error and scores 0.0.