---
dataset_info:
  features:
  - name: state
    dtype: string
  - name: turn
    dtype: string
  - name: negamax
    list: string
  - name: optimal
    dtype: string
  - name: stage
    dtype: string
  - name: outcome
    dtype: string
  - name: final_move
    dtype: int64
  - name: depth
    dtype: int64
  - name: assistant_threats
    dtype: string
  splits:
  - name: train
    num_bytes: 505934787
    num_examples: 2000000
  - name: test
    num_bytes: 19273018
    num_examples: 76145
  download_size: 183716944
  dataset_size: 525207805
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
license: apache-2.0
tags:
- connect4
- board_game
---


# Parsenal/c4_v2

The dataset contains about **2.076M positions**, with **76.1k positions reserved for the test split**.

## Overview

This dataset was created to provide rich supervision for Connect Four models.  
For each board position, the action value of every valid move is computed using a solver:

- Solver: https://github.com/ChristopheSteininger/c4.git

The core signal is the solver-produced **negamax list** over valid moves.  
Aside from threat annotations, most other targets can be derived directly from this list.

## Data generation

- Games were generated by solver play with **epsilon-greedy** action selection.
- This produced a distribution of positions from roughly **250k games**.
- Positions were shuffled.
- **76.1k positions** were selected for the test split.
- **Mirror duplicates were intentionally excluded**.

## Labels and supervision

For each position, the dataset provides supervision that can support several training objectives:

- **Optimal move**
- **Threat masks**
- **Per-move action values** for valid moves
- Signals that can be reduced to **position value** or related targets from the action-value / negamax outputs

### Threat definitions

<p align="left">
  <img src="threats.png" alt="Connect Four example position" width="300">
</p>
Threat annotations follow the definitions from **James D. Allen, _The Complete Book of Connect Four_**.

### Optimal move tie-breaking

When more than one move is optimal, the selected target move is chosen by the following rules:

1. Choose the **most central** optimal move.
2. If there is still a tie, choose the **leftmost** move.

## Intended use

This dataset was used to train **`granite_4_in_a_row`** by teaching the model to:

- output **threat masks**
- predict the **optimal move**

However, the dataset contains much richer supervision than that setup alone. It can also be used for:

- **value prediction**
- **action-value prediction**
- multi-task training over **threats + policy + value**

## Notes and limitations

- Because games are generated with **epsilon-greedy** play, the position distribution is not the same as pure optimal play or human play.
- **Mirrored positions are excluded**, so to try symmetry augmentation you should add it separately.

## Attribution

Solver used for annotation and game generation:

- https://github.com/ChristopheSteininger/c4.git

Threat definitions based on:

- James D. Allen, _The Complete Book of Connect Four_