tic-tac-toe

Environment ID: tic-tac-toe
Short description: Multi-turn Tic-Tac-Toe tool environment with configurable text or image board state.
Tags: multi-turn, tools, game, train, eval

Primary dataset(s): Synthetic repeated starting-board examples generated in load_environment.
Source links: None.
Split sizes: Controlled by num_examples and defaults to 100 examples.

Type: Multi-turn tool use.
Output format expectations: The model should call the exposed move tool with a board position from 0 to 8.
Rubric overview: win_reward returns 1.0 when X wins and 0.0 otherwise.

Run an image-board evaluation with default settings:

prime eval run tic-tac-toe

Run a text-only evaluation:

prime eval run tic-tac-toe -a '{"mode": "text"}'

Run an image-only evaluation explicitly:

prime eval run tic-tac-toe -a '{"mode": "image"}'

Configure model and sampling:

prime eval run tic-tac-toe -m gpt-4.1-mini -n 20 -r 3 -t 1024 -T 0.7 -a '{"mode": "text"}'

Use -a / --env-args to pass environment-specific configuration as a JSON object.

Arg	Type	Default	Description
`mode`	str	`"image"`	Board modality. Use `"text"` for ASCII boards, `"image"` for PNG board images, or `"mixed"` for alternating rows. Aliases include `"visual"` for image mode and `"ascii"` for text mode.
`num_examples`	int	`100`	Number of repeated starting-board examples to include. Must be at least 1.
`mixed_batch`	bool	`null`	Deprecated compatibility flag. `true` maps to `mode="mixed"`; `false` maps to `mode="image"`.

Metric	Meaning
`reward`	Main scalar reward; 1.0 when X wins, 0.0 otherwise.
`win_reward`	Raw binary win signal from the rubric.