chrisiyer/cardgames-sftdata-trimmed

--- pretty_name: Cardgames SFT Data (Trimmed) task_categories: - image-text-to-text language: - en size_categories: - 10K<n<100K configs: - config_name: blackjack data_files: - split: train path: blackjack/train-* - split: validation path: blackjack/validation-* - split: test path: blackjack/test-* - config_name: numberline data_files: - split: train path: numberline/train-* - split: validation path: numberline/validation-* - split: test path: numberline/test-* - config_name: ezpoints data_files: - split: train path: ezpoints/train-* - split: validation path: ezpoints/validation-* - split: test path: ezpoints/test-* - config_name: points24 data_files: - split: train path: points24/train-* - split: validation path: points24/validation-* - split: test path: points24/test-* --- # Cardgames SFT Data (Trimmed) ## Dataset Summary This dataset is a modified version of data published alongside *Fine-Tuning Large Vision-Language Models as Decision-Making Agents via Reinforcement Learning* ([website](https://rl4vlm.github.io), [data](https://huggingface.co/LEVI-Project/sft-data/tree/main)). If you use this data, please make sure to cite their work! It contains four card- and game-based vision-language decision tasks: 1. `numberline` 2. `ezpoints` 3. `points24` 4. `blackjack` Compared with the original published SFT data, this version simplifies prompts and target outputs so that each example contains only the image, prompt, and target output. In tasks that were originally structured as sequential action traces, we also collapse those traces into single supervised examples when appropriate. Each task is provided as its own dataset config with `train`, `validation`, and `test` splits, and each example contains the fields: - `id` - `image` - `prompt` - `output` The dataset is fully compatible with Hugging Face `datasets` and can be loaded directly with `load_dataset()`. Example: ```python from datasets import load_dataset dataset = load_dataset("chrisiyer/cardgames-sftdata-trimmed", "blackjack") train_split = dataset["train"] example = train_split[0] print(example["prompt"]) print(example["output"]) print(example["image"]) ``` ## Task Descriptions ### 1. Numberline Given a target number and a current number, presented in an image, the model must decide whether to move the current number up or down. This task is largely unmodified from the original version. The prompts and outputs were shortened to include just the image, prompt, and target output, excluding intermediate outputs such as target chain-of-thought text and other auxiliary fields. ### 2. EZPoints Given two cards that either add up to 12 or multiply to 12, the model must give a formula using their values that evaluates to 12, for example `3*4`. The prompts and outputs were shortened. We also eliminated the sequential design of the published SFT outputs, so that each trial is a single example in which the image is presented once and the full formula appears in the output. ### 3. Points24 Given four cards, the model must give a formula using their values that evaluates to 24. The prompts and outputs were shortened, and the sequential structure was eliminated. Note that models in the original paper performed poorly on this task. ### 4. Blackjack Given a dealer hand and a player hand, the model must decide whether to `hit` or `stand`. The prompts and outputs were shortened. The sequential trials were retained, for example when a trial following a `hit` contains the same cards plus one additional card, but every trial has a unique target label and can be treated as an independent supervised example. ## Source Data Original source data: - RL4VLM project website: [https://rl4vlm.github.io](https://rl4vlm.github.io) - Published SFT data: [https://huggingface.co/LEVI-Project/sft-data/tree/main](https://huggingface.co/LEVI-Project/sft-data/tree/main) ## Notes This dataset is intended as a simplified supervised fine-tuning version of the original task data for vision-language model training and continual-learning experiments.pretty_name: 卡牌游戏监督微调（Supervised Fine-Tuning，SFT）数据集（精简版） task_categories: - 图像-文本转文本 language: - 英语 size_categories: - 10000 < 样本量 < 100000 configs: - config_name: 二十一点（blackjack） data_files: - split: 训练集 path: blackjack/train-* - split: 验证集 path: blackjack/validation-* - split: 测试集 path: blackjack/test-* - config_name: 数轴任务（numberline） data_files: - split: 训练集 path: numberline/train-* - split: 验证集 path: numberline/validation-* - split: 测试集 path: numberline/test-* - config_name: 简易计分任务（ezpoints） data_files: - split: 训练集 path: ezpoints/train-* - split: 验证集 path: ezpoints/validation-* - split: 测试集 path: ezpoints/test-* - config_name: 二十四点任务（points24） data_files: - split: 训练集 path: points24/train-* - split: 验证集 path: points24/validation-* - split: 测试集 path: points24/test-* --- # 卡牌游戏监督微调数据集（精简版） ## 数据集概述 本数据集是《基于强化学习的大视觉语言模型作为决策AI智能体微调》一文配套发布数据的修改版本，相关链接如下：[项目官网](https://rl4vlm.github.io)、[原始数据集](https://huggingface.co/LEVI-Project/sft-data/tree/main)。 若使用本数据集，请务必引用原作者的研究工作。 本数据集包含四类基于卡牌与游戏的视觉语言决策任务： 1. 数轴任务（numberline） 2. 简易计分任务（ezpoints） 3. 二十四点任务（points24） 4. 二十一点（blackjack） 相较于原始发布的监督微调数据集，本版本简化了提示词与目标输出，每条样本仅包含图像、提示词与目标输出。对于原本采用序列动作轨迹结构的任务，我们在合适的场景下将轨迹压缩为单条监督学习样本。 每个任务均作为独立的数据集配置（config）提供，包含训练集、验证集与测试集划分，每条样本包含以下字段： - `id`：样本标识符 - `image`：图像数据 - `prompt`：提示词 - `output`：目标输出 本数据集完全兼容Hugging Face `datasets`库，可直接通过`load_dataset()`函数加载。 示例代码如下： python from datasets import load_dataset dataset = load_dataset("chrisiyer/cardgames-sftdata-trimmed", "blackjack") train_split = dataset["train"] example = train_split[0] print(example["prompt"]) print(example["output"]) print(example["image"]) ## 任务说明 ### 1. 数轴任务（numberline） 给定图像中展示的目标数字与当前数字，模型需判断将当前数字向上或向下移动。本任务基本保留原始版本的结构，仅简化了提示词与输出，仅保留图像、提示词与目标输出，移除了思维链文本等中间输出与其他辅助字段。 ### 2. 简易计分任务（ezpoints） 给定两张卡牌，其数值之和或乘积为12，模型需给出使用其数值计算得到12的算式，例如`3*4`。本任务简化了提示词与输出，同时移除了原始发布数据的序列结构，每条测试样本仅包含单次图像展示与完整输出算式。 ### 3. 二十四点任务（points24） 给定四张卡牌，模型需给出使用其数值计算得到24的算式。本任务简化了提示词与输出，同时移除了原始结构。需注意，原始论文中的模型在本任务上表现不佳。 ### 4. 二十一点（blackjack） 给定庄家手牌与玩家手牌，模型需判断执行`hit`（要牌）或`stand`（停牌）操作。本任务简化了提示词与输出，保留了序列测试样本结构——例如当某次测试执行要牌操作后，样本会包含原卡牌与新增的卡牌，但每条样本均拥有唯一的目标标签，可作为独立的监督学习样本处理。 ## 原始数据源 - RL4VLM项目官网：[https://rl4vlm.github.io](https://rl4vlm.github.io) - 发布的监督微调数据集：[https://huggingface.co/LEVI-Project/sft-data/tree/main](https://huggingface.co/LEVI-Project/sft-data/tree/main) ## 备注 本数据集为原始任务数据的精简版监督微调数据集，适用于视觉语言模型的训练与持续学习实验。