InternGo-LoGos

KO | EN

gitlite — search

#typescript #ai-agents #ai #dsh-plugin #deepseek-harness #open-source #claude-code #codex #cli #developer-tools #react #windows

★ 49 Open GitHub ↗

[NIPS 2025] Mixing Expert Knowledge: Bring Human Thoughts Back to The Game of Go. Our model is originally named InternThinker-Go, and called LoGos in our paper.

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

<div align="center">
<h1> Mixing Expert Knowledge: Bring Human Thoughts Back to The Game of Go
 </h1>
</div>

<p align="center">
We name our model <em><strong>LoGos</strong>,</em>  representing <em>"word"</em>  and
<em>"thought"</em> in Greek, aspiring to bring the uniquely human capacity of reasoning back to the ancient game of Go.
</p>

<p align="center">
📃 <a href="https://neurips.cc/virtual/2025/poster/117166" target="_blank">Paper</a> ｜ 🤗 <a href="https://huggingface.co/YichuanMa/LoGos-7B" target="_blank">LoGos-7B</a> ｜ 🤗 <a href="https://huggingface.co/datasets/YichuanMa/Expert-Go-SFT-100K" target="_blank">Datasets</a> ｜ 📧 <a href="mailto:mayichuan@pjlab.org.cn">Email</a>
</p>

## Overview
![LoGos Pipelin](https://raw.githubusercontent.com/Entarochuan/InternGo-LoGos/main/images/Go-Pipeline.png)
*Our methodology for integrating Go professional capabilities with LLMs' long COT reasoning abilities. After mixed cold start and GRPO training, our model successfully transfers the reasoning capabilities acquired from long CoT data to Go tasks. For a given query, the model correctly performs analysis, thinking, reasoning, and summarization, ultimately selecting a reasonable next move.*

## Demo
![LoGos Demo](https://raw.githubusercontent.com/Entarochuan/InternGo-LoGos/main/images/LoGos.gif)

## News

- **2025-09**: 🎉🎉Our paper is accepted by NIPS 2025.

- **2025-05**: We invite professional Go players 赵晨宇九段 and 战鹰二段 to play with InternThinker-Go, [see the live replay here](https://www.bilibili.com/video/BV1117wz3EHo/?vd_source=9d61fe3d4e46469f76cfd82a6da1ef35).

## **1. Collections**

### 1.1 Online Go Gameplay Platform

**[Click to play with our model online!](https://chat.intern-ai.org.cn/)**

### 1.2 Models and Datasets

You can find our latest models and datasets in the following two tables.

| Model Name | Description | Download Link | Size |
|------------|-------------|---------------|------|
| LoGos-7B | 7B parameter model for Go reasoning | [LoGos-7B](https://huggingface.co/YichuanMa/LoGos-7B) | 7B |

| Dataset Name | Description | Download Link | Size |
|--------------|-------------|---------------|------|
| Expert-Level Go Dataset | Expert Level Synthetic Go Dataset | [Expert-Go-SFT-100K](https://huggingface.co/datasets/YichuanMa/Expert-Go-SFT-100K) | 100K samples |
| GRPO-Dataset-1K | dataset used in reinforcement learning | [GRPO-1K](https://huggingface.co/datasets/YichuanMa/Go-GRPO-1K) | 1K samples |
| KataGo-Bench-1K | Evaluation Benchmark for Go | [KataGO-Bench-1k-eval.jsonl](KataGo-Bench-1K/GO_ELO/KataGO-1k-eval/eval-files/KataGO-Bench-1k-eval.jsonl) | 1K samples |
| LoGos-distillation-1K | Responses from LoGos 32B | [LoGos Rollout 1K](https://huggingface.co/datasets/YichuanMa/LoGos-Rollout-1K) | 1K samples |

## **2. KataGo-Bench-1K: A Go Benchmark for LLM**

A benchmark for evaluating LLMs on Go.

### Step1: Evaluation Setups

- **Setups for Go board rendering**

```bash
cd KataGo-Bench-1K
sudo apt update
sudo apt install nodejs npm
```

- **Setups for API call**

```bash
pip install openai
```

### Step2: Start Evaluation

```bash
python KataGo-Eval.py \
--api_base your_api_base_url \
--model_name your_model_name \
--api_key your_api_key \
# How the moves are placed in the query(basic: with space, numbered: with numerical order)
--input_mode numbered \
--num_threads num_workers \
# evaluation type: Reasoning_LM: for general LLMs, Addboard-KataGo-Eval: for LoGos series models
--task_type Addboard-KataGo-Eval \

```

### Step3: Review the Results
After execution, please see evaluation results in `./KataGO-1k-eval/eval-results`.

Specifically, for a given model, two result files will be generated in the `/KataGO-1k-eval/eval-results/{model_name}/{Y}{M}{D}_{H}{M}{S}`:

1. `eval_results.jsonl`: A JSONL file containing all generation results.

2. `summary.txt`: Summary for evaluation results.

## **3. Setups for Reinforcement Learning**

We are using [verl](https://github.com/volcengine/verl) for reinforcement learning, so the following changes should be applied on a stable verl environment. Notably, the experiments are conducted on `verl v0.2`.

### Step1: Reward Function

1. Move `RL_utils/Go_reward.py` into `verl/utils/reward_score`.

2. Add the following code to `verl/verl/utils/reward_score/__init__.py`

```python
if data_source == 'GO':
    from . import Go_reward
    res = Go_reward.compute_score(solution_str, ground_truth)
```

### Step2: Start Training

1. Prepare RL training data (download GRPO-Dataset-1K and put the `train.parquet` and `test.parquet` into `verl/data/GO`).

2. Start GRPO training, we recommand using 8xA800.

```bash
bash Go_train_demo.sh
```

## **4. Useful Toolkits**
Here are some useful tools in the `useful_tools` folder.

1. `Board_visualize.py`:  A tool for visualizing a given move list.

2. `goGame`: A tool for effective Go board rendering, with all Go rules considered(including capture, ko, ...). This tool is modified from [SabakiHQ/go-board](https://github.com/SabakiHQ/go-board), and can be considered a wrapper for Python.

Run the following commands to install neccessary requirements.
```bash
sudo apt update
sudo apt install nodejs npm
```
 
And see `Use_demo.py` for common usage.

## **5. Citation**

```bibtex
@misc{ma2026mixingexpertknowledgebring,
      title={Mixing Expert Knowledge: Bring Human Thoughts Back To the Game of Go}, 
      author={Yichuan Ma and Linyang Li and Yongkang Chen and Peiji Li and Jiasheng Ye and Qipeng Guo and Dahua Lin and Kai Chen},
      year={2026},
      eprint={2601.16447},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2601.16447}, 
}
```

## **6. License**

![Code License](https://img.shields.io/badge/Code%20License-Apache%202.0-green)
![Data License](https://img.shields.io/badge/Data%20License-CC%20By%20NC%204.0-orange)

**Usage and License Notices:**
The data and code are intended and licensed for research use only. License: Attribution-NonCommercial 4.0 International. It should abide by the policy of OpenAI: https://openai.com/policies/terms-of-use

## **7. Acknowledgements**

We would like to express our sincere gratitude to the following platforms and projects for their contributions to the community:

- **[verl](https://github.com/volcengine/verl)**: For the reinforcement learning training infrastructure
- **[KataGo](https://github.com/lightvector/KataGo)**: For the Go engine and evaluation tools
- **[Yike](https://home.yikeweiqi.com/#/game)**: For part of the Go professional datasets

InternGo-LoGos

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README