| --- |
| license: apache-2.0 |
| language: |
| - en |
| task_categories: |
| - text-classification |
| size_categories: |
| - 100K<n<1M |
| --- |
| # JADE-Data |
|
|
| The 160,017 training examples used for [JADE](https://huggingface.co/theunnecessarythings/JADE). |
|
|
| Each example presents a world or context and asks for a decision. The corpus covers |
| reasoning, language, retrieval, tool selection, stateful actions, music, and |
| probabilistic forecasting. Examples are generated from programs, simulators, |
| solvers, and structured world state. |
|
|
| ## Load the data |
|
|
| ```python |
| import json |
| from huggingface_hub import hf_hub_download |
| |
| path = hf_hub_download("theunnecessarythings/JADE-Data", |
| "train.jsonl", repo_type="dataset") |
| with open(path) as file: |
| for line in file: |
| example = json.loads(line) |
| # Use example["state"], example["question"], and example["label"]. |
| ``` |
|
|
| ## Format |
|
|
| `train.jsonl` contains 123,350 multiple-choice examples (`choice`) and 36,667 |
| probabilistic yes/no examples (`noul`). Each record includes its ID, family, suite, |
| state, question, target, and source metadata. Multiple-choice targets use option |
| IDs; probabilistic targets retain their probabilities. Related examples share a |
| family ID. |
|
|
| This release contains the training split. Evaluation and calibration partitions |
| are kept separately. All training examples are synthetic; Decision Index and |
| JevBench benchmark records are not part of this corpus. |
|
|
| Licensed under Apache-2.0. |
|
|
