TRACE / NOTICE.md
d4x1's picture taoist's picture
Duplicate from taoist/TRACE
fefcb3a
|
Raw History Blame Contribute Delete
4.89 kB

Notice: Yelp Open Dataset & License Boundaries

The TRACE benchmark builds on the Yelp Open Dataset. We do not redistribute Yelp-owned content. This file documents what is and isn't included, and why.

What is included in this release

Component Origin License
Dialogue structure (turns, actions, scenarios, personas, persona-grounded preferences) Generated by us CC BY 4.0
Dialogue text outside quotes Generated by us using LLMs CC BY 4.0
Per-turn grounding annotations (referenced_poi_ids, referenced_review_ids, acceptable_alternative_poi_ids, discovered_review_ids) Generated by us CC BY 4.0
quote_spans metadata (review IDs, integer offsets, edit operations) Generated by us CC BY 4.0
Difficulty labels, candidate pools, open-set pools Generated by us CC BY 4.0
Code (trace_kit/, scripts/, examples/) Written by us CC BY 4.0
Yelp business_id / review_id values Yelp Open Dataset Bare identifiers — factual reference, not subject to copyright
POI metadata (name, address, stars, categories, lat/lon, price range, attributes) Yelp Open Dataset Distributed under the academic-fair-use carve-out of §4.E (academic publication) — see below
Per-review metadata (stars, date) — IDs only Yelp Open Dataset Same as above

What is not included

Component Why
Verbatim review text Yelp ToS §4.A: cannot publicly display reviews.
LLM-generated quotations of reviews Even paraphrased quotes ≥90 fuzzy similarity are arguably derivative works of Yelp content (§5 Ownership). We replace them with [Q:N] placeholders backed by positional metadata.
Tip text, user account data Yelp ToS §4.A.
Bulk POI listings without provenance Would constitute "creating a competing local-listings DB" (§4.A).

To obtain review text, see scripts/reconstruct_kb.py — it pulls text from your own copy of yelp_academic_dataset_review.json (you must download it after accepting the Yelp ToS) and rebuilds both the KB and the dialogue text.

Quote-masking design

The quote_spans metadata is purposefully designed to contain no Yelp word content:

  • Exact verbatim quotes (≈86% of cases): stored as (review_id, [start, end]) — three integers + an ID.
  • Near-verbatim quotes (≈11% of cases): stored as (review_id, [start, end], edits). The edits list contains:
    • ins_p/rep_p: insert/replace using whitelisted punctuation only (. , ; : ! ? ' " - / ( ) [ ] { } whitespace). No alphanumeric content.
    • ins_r/rep_r: insert/replace by referencing another review span (rid, c, d). The actual word content is fetched from the user's hydrated Yelp data, never from our metadata.

Concretely, for any LLM-substituted alphanumeric content, the metadata stores only positional pointers into the corpus. A reader of our distribution who has not also accepted the Yelp ToS cannot reconstruct any Yelp word. This mirrors the "ship the IDs, hydrate locally" pattern used by every Yelp-derived academic benchmark.

The remaining ≈3% of LLM-quoted content (substantive paraphrase or non-quote regex artifacts) is left as plain text — these are demonstrably original LLM output that does not retrieve back to a single review span at ≥90 fuzzy similarity.

Yelp Dataset Terms of Use — key clauses (Feb 2021)

Authoritative source: https://www.yelp.com/dataset. The clauses below are summarized; the full agreement controls.

  • §3 License grant: royalty-free, non-exclusive, non-sublicensable, non-transferable, revocable, for non-commercial use only.
  • §3 Publication: prior to public presentation involving the Data and the Yelp brand name, results must be submitted to Yelp for review (5 business days).
  • §4.A: cannot publicly display the Data, especially reviews and other user generated content.
  • §4.B: no commercial use.
  • §4.E: cannot redistribute summaries or metrics — except disclosures necessary for academic publication.
  • §4.H: cannot rent, lease, sell, transfer, assign, or sublicense any part of the Data.
  • §10 Term: 12 months, terminable at any time; on termination must delete all copies.

How to comply when using TRACE

By downloading and using TRACE, you acknowledge that:

  1. You will obtain your own copy of the Yelp Open Dataset (and accept Yelp's ToS) before hydrating review text.
  2. You will use the resulting hydrated KB for non-commercial academic research only.
  3. You will not redistribute review text.
  4. If your institution has commercial interests, you will obtain a separate authorization from Yelp before using the data in production.

Reporting issues

  • Yelp ToS questions: dataset@yelp.com
  • TRACE issues: see the repository linked in dataset_info.json.