Downloads · 30 days
1
4% of all-time downloads
alexneakameni/Qwen3-4B-Instruct-2507-chess-grpo
Qwen3-4B-Instruct-2507-chess-grpo is a machine learning model from alexneakameni. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for transformers.
[](https://www.aicrowd.com/challenges/global-chess-challenge-2025)
Downloads · 30 days
1
4% of all-time downloads
All-time downloads
24
Public
Repo size
23.2 MB
Likes
0
Public
Click a slice to open those files.
.json14.2 MB · 51%
From the Hugging Face model README
This repository is the Submission template and Starter kit for the Global Chess Challenge! Clone the repository to compete now!
This repository contains:
Most chess players don't have regular access to a top coach. What they do have are their own games and a recurring question: "What should I have played here?" The Global Chess Challenge imagines a tool that looks at those positions, suggests a strong move, and explains the idea in simple language, so players can coach themselves using the games they already play.
This challenge asks you to build models that play legal chess moves and briefly explain their choices in natural language, while a world-class engine checks how well those moves hold up on the board. The challenge turns a familiar game into a testbed to see whether reasoning models can think clearly, play good moves, and talk about them in a way humans can follow.
The Global Chess Challenge asks participants to build a text-only chess agent that does two things at once: play a legal move and explain the idea behind it in simple language.
On each turn, your model receives a chess position as text and must respond with:
The environment verifies legality, evaluates move quality using Stockfish, and runs full games in tournaments to measure overall playing strength.
For every turn, your agent receives:
Your agent must return:
<think>...</think><uci_move>...</uci_move>Please follow the instructions in player_agents/README.md for instructions and examples on how to write your own chess agent for this competition.
Clone the repository recursively:
bash git clone --recursive [email protected]:aicrowd/global-chess-challenge-2025-starter-kit.git cd global-chess-challenge-2025-starter-kit
In case you didn't clone with --recursive, you can do the following
bash cd global-chess-challenge-2025-starter-kit git submodule update --init --recursive
Install competition specific dependencies
bash pip install -r requirements.txt
Before running local evaluation, you need to start either a vLLM server or a Flask server in a separate terminal from the player_agents directory.
cd player_agents
pip install vllm
bash run_vllm.sh
cd player_agents
# For random agent
python random_agent_flask_server.py
# OR for Stockfish agent
python stockfish_agent_flask_server.py
Keep this server running in the background while you run local evaluation.
Test your agent locally using python local_evaluation.py.
Note: Make sure you have started either the vLLM server or Flask server (see Running LLM locally) in a separate terminal before running local evaluation.
Accept the Challenge Rules on the main challenge page by clicking on the Participate button.
This guide walks you through the process of submitting your chess agent to the Global Chess Challenge 2025.
Before making a submission, ensure you have:
requirements.txt)First, authenticate with AIcrowd:
aicrowd login
You'll be prompted to enter your AIcrowd API key. You can find your API key at: https://www.aicrowd.com/participants/me
Your model must be hosted on Hugging Face. You can use:
Qwen/Qwen3-0.6B)If your model is private or gated, you need to grant AIcrowd access. See docs/huggingface-gated-models.md for detailed instructions.
Your prompt template should be a Jinja file that formats the chess position and legal moves for your model. Examples are available in the player_agents/ directory:
llm_agent_prompt_template.jinja - For general LLM agentssft_agent_prompt_template.jinja - For supervised fine-tuned agentsrandom_agent_prompt_template.jinja - Minimal template exampleEdit the aicrowd_submit.sh file with your submission details:
# Configuration variables
CHALLENGE="global-chess-challenge-2025"
HF_REPO="YOUR_HF_USERNAME/YOUR_MODEL_NAME" # e.g., "Qwen/Qwen3-0.6B"
HF_REPO_TAG="main" # or specific branch/tag
PROMPT_TEMPLATE="player_agents/YOUR_PROMPT_TEMPLATE.jinja"
global-chess-challenge-2025)username/model-name)main)If you're submitting a model that will run on AWS Neuron hardware (Trainium), make sure you pick the correct --neuron.model-type and review the supported --vllm.* flags you can pass to aicrowd submit-model.
See: docs/neuron-and-vllm-tuning.md
Once configured, run the submission script:
bash aicrowd_submit.sh
Or submit directly using the AIcrowd CLI:
aicrowd submit-model \
--challenge "global-chess-challenge-2025" \
--hf-repo "YOUR_HF_USERNAME/YOUR_MODEL_NAME" \
--hf-repo-tag "main" \
--prompt-template-path "player_agents/YOUR_PROMPT_TEMPLATE.jinja"
Games are played in round-robin tournaments with ACPL ratings determining the final rankings. Each move is also checked for legality and compared against Stockfish for calculating CPL.
Your agent must be self-contained and run without network access during evaluation. You can use Stockfish locally during training. During inference, we only run the LLM.
Best of Luck :tada: :tada: