chrisiyer/cardgames-sftdata-trimmed

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, data).

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:

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
• Published SFT data: 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.