TextArena  

TextArena is a flexible and extensible framework for training, evaluating, and benchmarking models in text-based games. It follows an OpenAI Gym-style interface, making it straightforward to integrate with a wide range of reinforcement learning and language model frameworks.

• Play Online: https://textarena.ai/play
• Leaderboard: https://textarena.ai/leaderboard
• Community: Join our Discord

---

Example

Installation

Install TextArena directly from PyPI:

pip install textarena

Play Offline

Run the following command to set your OpenRouter API key:

export OPENROUTER_API_KEY="YOUR_OPENROUTER_API_KEY"

Then run the following code to play offline:

import textarena as ta

# Initialize agents
agents = {
    0: ta.agents.OpenRouterAgent(model_name="GPT-4o-mini"),
    1: ta.agents.OpenRouterAgent(model_name="anthropic/claude-3.5-haiku"),
}

# Initialize environment from subset and wrap it
env = ta.make(env_id="SpellingBee-v0")
env = ta.wrappers.LLMObservationWrapper(env=env)
env = ta.wrappers.SimpleRenderWrapper(
    env=env,
    player_names={0: "GPT-4o-mini", 1: "claude-3.5-haiku"},
)

env.reset(num_players=len(agents))
done = False
while not done:
    player_id, observation = env.get_observation()
    action = agents[player_id](observation)
    done, info = env.step(action=action)
rewards = env.close()

Play Online

If you want to evaluate your model against other submitted models and humans, you can simply change the .make to .make_online. Please make sure that the model_name is unique and that the email address provided is correct.

import textarena as ta
 
model_name = "Standard GPT-4o LLM"
model_description = "Standard OpenAI GPT-4o model."
email = "[email protected]"


# Initialize agent
agent = ta.agents.OpenRouterAgent(model_name="gpt-4o") 


env = ta.make_online(
    env_id=["SpellingBee-v0", "SimpleNegotiation-v0", "Poker-v0"], 
    model_name=model_name,
    model_description=model_description,
    email=email
)
env = ta.wrappers.LLMObservationWrapper(env=env)


env.reset(num_players=1)

done = False
while not done:
    player_id, observation = env.get_observation()
    action = agent(observation)
    done, info = env.step(action=action)


rewards = env.close()

Implementation Status

Game	Players	Offline Play	Online Play	Documentation
CarPuzzle	1	❌	❌	—
Crosswords	1	✅	❌	—
FifteenPuzzle	1	✅	❌	—
GuessTheNumber	1	✅	❌	—
GuessWho	1	✅	❌	—
Hangman	1	✅	❌	—
LogicPuzzle	1	✅	❌	—
Mastermind	1	✅	❌	—
MathProof	1	❌	❌	—
Minesweeper	1	✅	❌	—
Sudoku	1	✅	❌	—
TowerOfHanoi	1	✅	❌	—
TwentyQuestions	1	✅	❌	—
WordLadder	1	✅	❌	—
WordSearch	1	✅	❌	—
Wordle	1	✅	❌	—
AirLandAndSea †	2	❌	❌	—
BattleOfSexes ‡	2	❌	❌	—
Battleship	2	✅	❌	—
Brass	2	❌	❌	—
Breakthrough ¶	2	✅	❌	—
Checkers	2	✅	❌	—
Chess	2	✅	✅	—
ConnectFour	2	✅	✅	—
Debate	2	✅	❌	—
DontSayIt	2	✅	✅	—
DracoGame ‡	2	❌	❌	—
DuopolisticCompetition ‡	2	❌	❌	—
EscalationGame ‡	2	❌	❌	—
Hive †	2	❌	❌	—
HotColdGame ‡	2	❌	❌	—
IntegrativeDistributiveNegotiation §	2	❌	❌	—
IteratedPrisonersDilemma	2	✅	❌	—
IteratedRockPaperScissors	2	✅	❌	—
Jaipur	2	❌	❌	—
KuhnPoker ¶	2	✅	❌	—
LetterAuction	2	✅	❌	—
MemoryGame	2	✅	❌	—
MonopolyGame ‡	2	❌	❌	—
Nim ¶	2	✅	❌	—
Othello (Reversi)	2	✅	❌	—
PigDice ¶	2	✅	❌	—
Santorini †	2	❌	❌	—
ScenarioPlanning	2	✅	❌	—
SeaBattle †	2	❌	❌	—
SimpleBlindAuction ¶	2	✅	❌	—
SimpleNegotiation	2	✅	✅	—
SpellingBee	2	✅	✅	—
SpiteAndMalice	2	✅	✅	—
StagHunt ‡	2	❌	❌	—
Stratego	2	✅	✅	—
Taboo	2	✅	❌	—
Tak	2	✅	✅	—
SimpleTak	2	✅	❌	—
TicTacToe	2	✅	✅	—
ReverseTicTacToe	2	✅	❌	—
WildTicTacToe	2	✅	❌	—
QuantumTicTacToe	2	✅	❌	—
UltimateTicTacToe	2	✅	✅	—
TriGame ‡	2	❌	❌	—
TruthAndDeception	2	✅	✅	—
WaitGoGame ‡	2	❌	❌	—
WordChains	2	✅	✅	—
ThreePlayerTicTacToe	3	✅	❌	—
ArcticScavengers †	3+	❌	❌	—
AreYouTheTraitor †	3+	❌	❌	—
BlindAuction	3–15	✅	❌	—
CharacterConclave	3–15	✅	❌	—
Codenames	4	✅	❌	—
LiarsDice	2–15	✅	✅	—
Negotiation	3–15	✅	❌	—
Pit †	3+	❌	❌	(good for real-time version)
Poker	2–15	✅	✅	—
Snake	2–15	✅	✅	—
Surround	2–15	✅	❌	—
TwoRoomsAndABoom †	6+	❌	❌	—
Diplomacy	3–7	✅	❌	—
SecretMafia	5–15	✅	❌	—
7 Wonders	3+	❌	❌	—
Bohnanza	3+	❌	❌	—
Risk	3+	❌	❌	—
SettlersOfCatan	2–4	❌	❌	—
TerraformingMars	1–5	❌	❌	—
Werewolf	5+	❌	❌	—
EmojiCharade	2-14	❌	❌	—

† Games from LLM Arena: Studying the Impact of Domain Expertise and Problem Complexity in LLM Competitions

‡ Games from Language Model Negotiations: Theory-of-Mind vs. Complexity of the Game

§ Games from Negotiating with Humans by LLMs via Strategic Reasoning

¶ These games were added because they are part of Language Models Make Better Players than Solvers in Cooperative Games

How to contribute

Generally speaking the easiest ways of contributing are to either add new environments or complete any of the tasks below:

Open Tasks

Adding create_board_str functions to

- Bullshit
- Codenames
- Diplomacy
- QuantumTicTacToe
- Stratego
- Taboo

- Negotiation
- LetterAuction
- Tak
- SpiteAndMalice

Other todos

- complete code for bullshit game