TRAVERSE

TRAVERSE (named for the practitioner’s term for contesting a rejection) is a corpus of linked rejection–response exchanges from United States patent prosecution. Each row is one claim-level rejection instance: a statutory rejection of one claim at one office action, paired with the applicant’s answer (the amended claim and remarks) and labeled by whether the rejection was sustained on the same grounds at the next office action.

TRAVERSE supports prediction of rejection outcomes using the examiner’s rejection, the applicant’s response, and information available at the time of examination.

Composition

• 2,102,408 labeled instances across 148,781 exchanges in 95,683 applications
• Coverage: applications classified in the G06F area of the Cooperative Patent Classification (software and digital data processing). Every instance’s office action was mailed on or after October 1, 2017, and its response is dated on or after February 1, 2019, so all responses and outcomes postdate the 2019 Revised Patent Subject Matter Eligibility Guidance (effective January 7, 2019). A small share of § 101 instances (1.8 percent) answers a rejection mailed before that effective date.
• Applications were selected by CPC classification rather than random sampling, preserving examiner and art-unit histories within G06F

Ground	Instances	Sustain rate
§ 101	567,843	78.4%
§ 102	264,569	14.3%
§ 103	931,765	16.6%
§ 112(a)	55,972	31.4%
§ 112(b)	282,259	40.2%

Split: train 1,658,961 / test 443,447, by response date (test = responses on or after 2024-05-31). The split is temporal; keep it that way to avoid leakage through examiner histories.

Fields

Per instance: identifiers (application, office actions n and n+1, claim version and number), statute_fine (the ground), label (1 = the rejection is sustained on the same grounds at the next office action under the ground’s comparison rule; 0 = it is not; for §§ 102 and 103 a 0 can coexist with a new rejection under the same statute — see Label rules), split, dates (office action, response), examiner_id and art_unit, as-of examiner and art-unit sustainment statistics (computed only from outcomes resolved before the office action’s calendar quarter, to prevent temporal leakage), and four texts: the examiner’s rejection text for the ground, the rejected claim, the amended claim, and the applicant’s remarks.

Label rules (summary)

§ 101: sustained iff the next office action again rejects the claim under § 101. § 112(a)/(b): sustained iff the same subsection persists (per-subsection membership). § 102: sustained iff any originally cited reference is cited again. § 103: sustained iff the next rejection’s reference set is a subset of the original.

Every instance has a next office action. A label of 0 (not sustained under the same ground) means the original basis did not carry forward under the rule above, not that the claim escaped rejection. The examiner may re-reject under the same statute on a new basis: in the source corpus, 68.2 percent of § 103 instances labeled 0 are rejected again under § 103 at the next office action on a different reference combination, and 23.1 percent of § 102 instances labeled 0 are re-rejected under § 102. The target measures whether the original rejection basis persists under the label rules above.

Intended use and limits

TRAVERSE is for method development and comparison on the rejection-persistence task. It is a subset of a larger research corpus; models trained on this subset will not reproduce the AUCs reported in the source study, which depend on the full corpus’s examiner-history volume. Rows derive from public USPTO records; older documents were produced by OCR and were filtered on per-character confidence, but residual recognition errors remain. All four text fields are populated; 31 exchanges whose remarks document is not recoverable from the source corpus were excluded. TRAVERSE enforces its date windows at the instance level; the source corpus is broader, and a small number of rows (0.3%) whose recorded dates are internally inconsistent in the source records were also excluded. Labels are constructed, not adjudicated: they encode the comparison rules above applied to structured USPTO citation and rejection data.

Provenance and license

The dataset compiles public patent-prosecution records distributed by the United States Patent and Trademark Office through its bulk data products and APIs. Examiner-authored material is a work of the United States government (17 U.S.C. § 105). Applicant-filed material (claims, amendments, remarks) is reproduced as it appears in the public prosecution record, the basis on which the USPTO publishes and disseminates it (see 37 C.F.R. § 1.71(d)–(e)). The compilation itself (the selection, linking, labels, and statistics) is dedicated to the public domain under CC0 1.0 Universal. The dedication covers everything the author has rights in; underlying documents carry whatever rights, if any, survive their inclusion in the public record. If you use TRAVERSE, please cite:

Howell, D. S. (2026). Prediction of office actions for patent applications [Doctor of Engineering praxis, The George Washington University].

Contact

Dataset questions: open a discussion on this repository.