README.md

metadata

license: mit
configs:
  - config_name: level1
    data_files:
      - split: train
        path: level1/train.parquet
      - split: test
        path: level1/test.parquet
  - config_name: level2
    data_files:
      - split: train
        path: level2/train.parquet
      - split: test
        path: level2/test.parquet
  - config_name: level3
    data_files:
      - split: train
        path: level3/train.parquet
      - split: test
        path: level3/test.parquet
  - config_name: level4
    data_files:
      - split: train
        path: level4/train.parquet
      - split: test
        path: level4/test.parquet
  - config_name: level5
    data_files:
      - split: train
        path: level5/train.parquet
      - split: test
        path: level5/test.parquet
  - config_name: level6
    data_files:
      - split: train
        path: level6/train.parquet
      - split: test
        path: level6/test.parquet
  - config_name: level7
    data_files:
      - split: train
        path: level7/train.parquet
      - split: test
        path: level7/test.parquet
  - config_name: levelX
    data_files:
      - split: train
        path: levelX/train.parquet
      - split: test
        path: levelX/test.parquet
pretty_name: BaLLMatro
task_categories:
  - text-generation
size_categories:
  - 1K<n<10K

A challenging task for LLMs in which they need to create high-scoring Ballatro-like hands.

Dataset Details

Dataset Description

BaLLMatro is a portmanteu of "LLM" (Large Language Model) and "Ballatro", the critically acclaimed videogame. Inspired by the layers of complexity of such game, this dataset tries to test the ability of LLMs in finding high-scoring "augmented" poker hands, under increasingly complex scoring rules. Thus, the objective of the dataset is to find the generalization abilities of LLMs, in a task where both humans and AI models can measure their performance.

Each split of the dataset represents an increasing level of difficulty. For further details on the task and how each level was generated, refer to the BaLLMatro project github page.

• Curated by: Álvaro Barbero Jiménez
• License: MIT

Dataset Structure

Every level contains a train and a test fold. A fair use of this dataset involves using the test folds only for testing the LLM peformance, which means:

• Not using test data to fine-tune the model in any way.
• Not using test data as few-shot examples, or any other kind of in-context learning or prompting approach that makes use of test examples or informacion about the distribution or nature of test examples.
• Not using the source code of the BaLLMatro repository to provide context, prompt, or train the LLM in any way. Only the section "The rules of BaLLMatro" can be used as a prompt to explain the LLM the rules of the game, but making use of source code or a list of jokers is not considered a fair use.

At all levels the following dataset columns are present:

• input: list of cards that must be used as input to the LLM. This is the only field that can be used as input to the LLM.
• output: optimal play, which can be used to compare against the output produced by the LLM being evaluated.
• score: score attained by the optimal play.
• hand: poker hand used in the optimal play.
• chips: number of chips obtained in the optimal play.
• chips: multiplier obtained in the optimal play.
• remaining: inpout cards not used in the optimal play.