Typed Decisions, test split
===========================

The file test.jsonl.gz in this directory is third-party data. It is not part of
tool-eval-bench's own MIT-licensed code, and its authors do not endorse this
project.

Dataset:   Typed Decisions (LocalLLaMA/typed-decisions)
Publisher: the LocalLLaMA organization on Hugging Face. The dataset card names
           no individual authors.
Source:    https://huggingface.co/datasets/LocalLLaMA/typed-decisions
Revision:  e039ebffcc280174dd354227424fb2b249f191de
File:      all/test-00000-of-00001.parquet (the test split, 400 cases)
License:   Apache License 2.0, as declared on the dataset card. The full text is
           in LICENSE in this directory.

Changes from the source, made by scripts/vendor_typed_decisions.py:

- Converted from Parquet to gzipped JSON Lines.
- Kept the id, workflow, state, questions and gold columns and dropped the rest
  (split, factors, label_agreement, n_questions).
- Decoded the state, questions and gold columns from JSON strings into JSON
  objects.

No row was added, removed, reordered or edited. manifest.json records the
sha256 of the source file, which the script checks before converting, and of
the converted data, which tool-eval-bench checks before every run.

The train split is deliberately not included.
