Szeweq/polish-decisions

Polish Decisions

Polish-language structured questions with a context state, candidate options, and a numeric target vector. The repository contains 170 records across train and eval splits. The dataset is intended to support training and evaluation of System One models, including Jev and Laya. Its Polish-language examples aim to address challenges in developing and training these models for Polish-language use.

Dataset summary

Split	Records	File
train	100	train.jsonl
eval	70	eval.jsonl

Each file is JSON Lines: one JSON object per line. Every record has the same five top-level fields. options and target have equal lengths in the supplied records; positions correspond.

Field	Stored shape	Description
question	string	Polish-language question or prompt.
state	string, object, or array	Context supplied with the question. Its JSON type varies between records.
options	array of strings	Candidate options; supplied records contain 2–5 options.
target	array of numbers	Numeric values aligned positionally with options.
type	string	Record label. Observed values are choice, score, and noul.

Target values

The files do not include a formal target specification. In the supplied records, choice targets are one-hot vectors, score targets are non-negative values that sum to approximately 1, and noul targets are binary vectors with one or more positive entries. The intended interpretation of noul and the distinction between these record types are not documented. Treat these descriptions as observed data patterns, not as a complete task definition.

Load the JSONL files

This standard-library example preserves the original JSON types, including the heterogeneous values in state:

import json
from pathlib import Path


def read_jsonl(path):
    with Path(path).open(encoding="utf-8") as file:
        return [json.loads(line) for line in file if line.strip()]


train = read_jsonl("train.jsonl")
evaluation = read_jsonl("eval.jsonl")

Tabular loaders may need an explicit normalization for state, because its values are not all the same JSON type. The Hub split configuration names the supplied evaluation split eval; no validation protocol is documented.

Synthetic examples

I disclose as the dataset author that some examples were generated synthetically with AI. The JSONL files do not identify which records are synthetic or report their proportion, the generation system, or the review process. Account for this when designing training or evaluation experiments.

Intended use

Use this dataset to explore training and evaluating System One models such as Jev and Laya with Polish-language examples. It is not validated guidance for real-world decisions or a benchmark with a documented scoring protocol.

Limitations

• Original sources, example-level provenance, collection method, and annotation process are not documented in the repository.
• The meanings of the type labels and target values are only partly inferable from their observed shapes; no metric or reference evaluation procedure is supplied.
• The split construction and any relationship between training and evaluation examples are undocumented.
• The dataset is small: the evaluation split contains 70 records. Results may be sensitive to individual examples and should not be generalized without further validation.
• The repository does not document a privacy review, consent process, or personal-data handling. Review the records and their source before using or redistributing them.

License and rights

This dataset is dual-licensed. You may use it under either the Apache License 2.0 or the MIT License. The terms of the license you choose govern your use, modification, and redistribution of the dataset.

Citation

No publication or formal citation is provided with the dataset. If referencing this repository, cite the dataset page and the specific revision used:

https://huggingface.co/datasets/Szeweq/polish-decisions