widebluesky/wbs-embedding-demo

Independent task dataset v1

This private dataset packages the synthetic English fixtures used by the trained embedding demo model. It includes its own training records and a self-contained copy of the corpus, queries and relevance judgments needed for evaluation. Loading and training do not require another dataset repository or model task. Original project-native files are preserved byte-for-byte under source/; data/ contains explicit Datasets JSONL loading views with unchanged records.

These fixtures were authored in advai-notebooks, with no customer data or external benchmark source. They verify engineering contracts, not domain quality.

Configurations and split counts

Configuration	Split counts
queries	train: 12, validation: 3, test: 3
corpus	corpus: 12
qrels	train: 144, validation: 36, test: 36
embedding_training	train: 24

The default configuration is embedding_training.

Schema and evaluation semantics

• queries: query_id, text, group_id, split.
• corpus: doc_id, text, group_id; the corpus is shared across query evaluation splits within this dataset.
• qrels: query_id, doc_id, relevance (0 to 3), assigned to the referenced query split.
• embedding_training: query_id, positive_id, negative_ids; 24 contrastive training examples.

There are 12 documents, 18 queries and 216 synthetic judgments. Training views contain only train queries; validation/test query IDs remain in the separate benchmark views. Query groups and identical queries do not cross splits. Positive document source groups are isolated by query split. Negative labels are checked against judgments.

Ranker training pairs and candidates are excluded from this dataset.

Loading

from datasets import load_dataset

samples = load_dataset("widebluesky/wbs-embedding-demo", "embedding_training", revision="demo-v1", token=True)
queries = load_dataset("widebluesky/wbs-embedding-demo", "queries", revision="demo-v1", token=True)

Use an existing Hugging Face login; do not embed credentials. Pin the dataset commit SHA for reproducibility. For project-native training, download a dataset snapshot and point task configuration paths at source/corpus.jsonl, source/queries.jsonl, source/qrels.jsonl and source/embedding_samples.json.

Trained model and provenance

Corresponding model: widebluesky/wbs-embedding-demo.

Canonical benchmark fingerprint: f96589e88e79ce684f3c8733ae1a3c98a8f028459f6e82e3477efeb2867b2a19.

Training sample fingerprint: cebc151f0e9c04eae8c92a1d79a8c3ae6771b1a81fb7484e7a13a34555ee9173.

Project validators verify formats, IDs, references, splits and training/candidate rules. Fingerprints match the saved records from the run that produced the published model. Every configuration is reloaded locally and from a pinned Hub commit and all records are compared before updating model associations. Source hashes, view fingerprints and published model weight hashes are recorded in dataset_manifest.json. No weights, optimizer state, logs, credentials or machine-specific paths are included.

Migration

This task dataset supersedes the shared repository for this model's active association. The original shared dataset remains available at its historical revision: widebluesky/wbs-retrieval-demo. The same small benchmark records are intentionally copied into each task dataset; training records and dataset versioning are maintained independently.