---
license: unknown
language:
  - en
tags:
  - gomoku
  - vision-language
  - vlm
  - lora
  - synthetic-data
task_categories:
  - question-answering
  - image-text-to-text
pretty_name: "Gomoku VLM Dataset (LoRA finetuning)"
configs:
  - config_name: visual
    data_files:
      - split: train
        path: "visual/visual_train.parquet"
      - split: validation
        path: "visual/visual_eval.parquet"

  - config_name: strategy
    data_files:
      - split: train
        path: "strategy/strategy_train.parquet"
      - split: validation
        path: "strategy/strategy_eval.parquet"

  - config_name: visual_curriculum
    data_files:
      - split: train
        path:
          - "curriculum/step_1.parquet"
          - "curriculum/step_2.parquet"
          - "curriculum/step_3.parquet"
          - "curriculum/step_4.parquet"

  - config_name: test
    data_files:
      - split: test
        path: "test/combined.parquet"
---

# Gomoku VLM Dataset (LoRA finetuning)

This repository contains a synthetic, image-grounded instruction dataset for training and evaluating **vision-language models (VLMs)** on **Gomoku (15×15)**.  
The dataset is designed for **LoRA finetuning** of image-text-to-text vision-language models on **two complementary capabilities**:

- **Visual**  
  Tasks where the model must read the board image and produce a structured answer about the current position.  
  This includes purely perceptual objectives (cell classification, counting) and also *visually grounded reasoning* such as run/line detection, matrix reconstruction, end-state recognition, and yes/no tactical assessments that can be decided from the current snapshot (e.g., “immediate win exists”, “opponent threatens immediate win”).  
  
- **Curriculum:** Curriculum-learning variant for *visual skills*: the training data is split into four steps that progressively move from simpler to more complex board states and visually grounded objectives (e.g., from basic cell/count tasks toward more advanced structured board understanding)

- **Strategy / policy (action selection)**  
  Tasks that require choosing an action (e.g., best move / win-in-1 move selection) and decision-making that approximates a bot’s policy.

Each example includes:
1) a rendered board image,  
2) a natural-language question, and  
3) one or more valid ground-truth answers (string list).

---

## Dataset structure

This dataset is organized into multiple Hugging Face **configs** that mirror the repository folders:

## Configs

- **`visual`**  
  Perception-focused questions (board reading, counting, localization, etc.).  
  Splits:
  - `train` → `visual/train/*.parquet`
  - `validation` → `visual/eval/*.parquet`

- **`strategy`**  
  Tactical / strategic questions (e.g., win-in-1 style tasks, move selection based on bot-policy).  
  Splits:
  - `train` → `strategy/train/*.parquet`
  - `validation` → `strategy/eval/*.parquet`

- **`visual_curriculum`**  
  Step-wise curriculum training data as **four growing steps**:
  - `curriculum/step_1.parquet`
  - `curriculum/step_2.parquet`
  - `curriculum/step_3.parquet`
  - `curriculum/step_4.parquet`

- **`test`**  
  Test Dataset:
  - `test` → `test/combined.parquet`

## Downloading the dataset locally

### Make sure the hf CLI is installed
```bash
curl -LsSf https://hf.co/cli/install.sh | bash
```

### Source bashrc
```bash
source ~/.bashrc
```

### Download Dataset
```bash
hf download eganscha/gomoku_vlm_ds --repo-type=dataset --local-dir ./gomoku_vlm_ds
```