Datasets:

---

LG-AI-Research

/

PANORAMA

like 3

Follow

LG AI Research 17

Hi PANORAMA team — thank you for releasing this dataset. The decision-trail framing and the three sequential benchmarks (PAR4PC / PI4PC / NOC4PC) are genuinely valuable, and I've been building a NOC4PC baseline on top of it.

While doing so I found that a subset of records' gold labels cannot be derived from the provided prior_art_specifications, because the reference the examiner actually relied on is non-patent literature (NPL) that the dataset does not include.

Root cause (construction)

Per §2.2.1, cited references are collected by matching patent-number patterns and using patent_client:

"We use regular expressions to identify patent numbers in the format (e.g., Patent Application No. US 2025/1234567; Patent No. US 12345678) and the patent_client Library to extract the specification, abstract, and claims of the cited patents."

NPL references (journal/conference papers) cited in the Non-Final Rejection are therefore never collected — even when they are the determinative reference for a §102/§103 rejection.

Concrete, reproducible example

validation, application_number = 16535912, claim_number = 1, gold code = 102:

• The gold reason attributes the anticipating disclosure to a reference cited as "Biegi", with journal-style locators — e.g. "...encoding ... as an n-dimensional vector (page 216 section 4.1 left column topic-frequency vector eu ...)". The page / section / column form is the convention for a non-patent (journal) document.
• But prior_art_specifications for this record contains a single, different document: patent US 20170091303 ("Client-Side Web Usage Data Collection").
• That patent does not disclose the claim's n-dimensional vector / reduced-dimension vector / cluster limitations, and the dataset's own rationale does not claim it does — it attributes them to "Biegi". So a model given only prior_art_specifications cannot reach the §102 label for this record; the cited reference simply isn't there.

Scale (conservative lower bound)

Counting §102/§103 records whose reason contains explicit NPL-style locators (page N / section N.N / left|right column):

This is a lower bound — it only catches rationales that include an explicit page/section/column locator. Rationales that name an NPL reference by author surname without a locator are not counted here, so the true number of records whose cited reference is absent is likely higher.

Why it matters

For these records, NOC4PC is effectively ill-posed: the determinative reference is not in the model's input, so the §102/§103 gold label is unobtainable from prior_art_specifications. Aggregate NOC4PC accuracy will therefore understate model capability, since a fraction of the labels cannot be reached from the provided context. (The first sequential task, PAR4PC, may be affected analogously when the examiner-cited document is NPL — though I have only verified NOC4PC.)

§4 (Limitations) discusses scale, USPTO-only scope, and the absence of allowance rationales, but does not mention NPL exclusion or missing cited references.

Suggestions (any subset)

1. Document the NPL exclusion explicitly in the dataset card / Limitations.
2. Add a per-record flag indicating whether all examiner-cited references are present in prior_art_specifications, so users can evaluate on the reference-complete subset.
3. Optionally drop NPL-only §102/§103 records from NOC4PC/PAR4PC, or include the NPL text where licensing/availability permits.

Happy to share the small script I used to detect NPL-cited rationales (and a reference-grounding overlap check) if it would be useful. Thanks again for the work — this is a great resource and I'd love to use it on the most rigorous footing possible.