widebluesky/wbs-decision-demo

Demo dataset v1

Small, deliberately synthetic English fixtures used to train and evaluate the corresponding tiny demo models from advai-notebooks. The records were authored as repository fixtures, with no external benchmark or customer data source. This dataset verifies engineering paths and data contracts; it does not establish domain quality or independent generalization performance.

Original source files are preserved byte-for-byte under source/. JSONL views under data/ expose explicit subsets and splits through the Hugging Face Datasets library. The views preserve original values and ordering within each split.

Configurations and splits

Configuration	Split counts
triage	train: 12, validation: 4, calibration: 10, test: 4

Schema and split semantics

Each record contains id, group_id, split, task, state and gold. source/tasks.json supplies questions, option definitions and output types. The triage task has department (choice), urgency (score) and churn risk (noul). Each of the 30 message records has three labels, for 90 individual decisions. The 10 calibration records are separate from training, validation and test and are used only to fit temperature scaling. Groups and identical states do not cross splits.

Loading

from datasets import load_dataset

dataset = load_dataset("widebluesky/wbs-decision-demo", "triage", revision="demo-v1", token=True)
print(dataset)

These repositories are private. Use your existing Hugging Face login for access; do not embed credentials in code or dataset files. Pin the dataset commit SHA for reproducibility.

For project-native file loading, download the dataset snapshot and point configuration paths at its original source/ files. The derived subsets are additional loading views.

Corresponding trained models

• widebluesky/wbs-decision-demo: verified training data fingerprint 2ed63814602e43b30c66678574f4808d06b3fd01e829eab4c06f039d16be4c11.

Validation and provenance

Project validators checked IDs, schemas, references, labels, split groups and negative/candidate rules. Dataset fingerprints match saved records from the runs that produced the published model weights. The publication process reloads every subset locally and from a pinned Hub commit and compares all records. Source file hashes, subset record fingerprints and model weight hashes are recorded in dataset_manifest.json. Model weights, training state, logs and machine-specific paths are not included.

Source implementation commit: f43a44e2431bae9d0451b445cc3bde5f1e2b8493.