yuebanlaosiji/findable-lite

Findable Lite

Findable Lite is a public development set for implementing and debugging catalog-discovery agents, local fixture tools, and deterministic scoring contracts. Public tasks and rubrics may be absorbed into model training data. Do not use this dataset for formal ranking or leaderboard claims.

Identity

• Dataset: findable-lite@0.3.0
• Contract: v1-draft
• Dataset content root: sha256:2ec0cb867e40e3961be135701c09c7d53bc3891869eb63300cc6f8cb535005bf
• Catalog content root: sha256:f6cde48877cdbb5076082003013cb7d405c98a9939862ca1bd1e7485fa272a53

The dataset content root is locator-independent. It does not include repository URLs, branch names, revisions, artifact paths, this card, or its own root value.

Coverage

Unit	Count
Tasks	204
Evaluation states	340
Products	15
Variants	45
Offers	45
Categories	3

The catalog covers digital accessories, home appliances, and outdoor gear. It is intentionally bounded and does not represent production commerce diversity.

Each of the six public capability groups contains 34 tasks:

Capability	Tasks
Known-product variant selection	34
Complete hard-condition search	34
Scenario inference and trade-off	34
Ask when necessary information is missing	34
Dynamic multi-turn discovery	34
Revise conditions after the user changes their mind	34

Loading

This is a no-script Hugging Face dataset. Each config is one Parquet file and preserves nested arrays as native lists or lists of structs.

from datasets import load_dataset

tasks = load_dataset(".", "tasks")
flows = load_dataset(".", "flows")
evaluations = load_dataset(".", "evaluations")
catalog = load_dataset(".", "catalog")

Condition and fact values are heterogeneous in the canonical model, while Parquet has no union column. They are encoded as value_json plus value_type (string, number, boolean, array, object, or null) inside the corresponding nested struct. The Findable loader validates the tag and reconstructs the canonical value before scoring.

Known Rubric Gaps

1. TIGHTEN has no offer priced exactly at its retired budget value, so an isolated stale-fact negative requires synthetic evidence.
2. RELAX's terminal state does not declare stale_condition_rejection; stale-check conformance requires an explicitly recorded rubric override.

Synthetic Data And Real Names

Tasks, descriptions, product IDs, variants, offers, prices, inventory, options, and surrounding catalog facts are synthetic. Some product titles use real brand and model names solely to identify the referenced products. This is nominative fair use and does not imply affiliation, endorsement, or that the synthetic offers describe real commercial listings.

Machine-readable field provenance is in manifest.json#/provenance/fieldProvenance. Product-title fields are marked nominative_reference; synthetic fields are marked synthetic.

Licenses

This dataset is licensed under CC-BY-4.0. The Findable code repository is separately licensed under Apache-2.0. Neither license replaces the other.

Immutable Cross-References

• Code and contract: https://github.com/hwy-deeplumen/findable/commit/16c1834d43be277e32b760f12b7ff85602c2554c
• Dataset revision: https://huggingface.co/datasets/yuebanlaosiji/findable-lite/commit/b6f575625e7aa7fd5a8769e6971a499912a8d01e

These references identify the immutable implementation and dataset baselines from which the local 0.3.0 semantic expansion was authored. Do not use a moving default branch as a release identity.