LexPipe/time-entries-and-phases

---

Litigation Billable Hours: Time-Entries and Phases Dataset

AI-ready, real-world time-entries mapped to litigation phases.

A datset of hours billed to 31 different litigation matters, with the names of the timekeeper, narrative description, duration, date, and other meta-information.

Why release this data?

• Establish an open-source baseline. Time entries are typically confidential. Open data accelerates research.
• Public provenance. Every row comes from an exhibit to a Motion for Attorneys’ Fees, these are public time entries.
• Encourage reproducible benchmarks. A common dataset lets the community compare narrative classifiers and cost models.
• Teach. Useful for classroom demos on legal operations data.
• Down-stream Products Seed data for proof-of-concepts.

Brought to you by

Created by LexPipe, which sourced time entries from public records. LexPipe found examples of full attorney invoices as exhibits to Motion for Attoneys Fees. The result is the first set of public royalty free time entries for research and commerce.

Whether you’re a law firm, ALSP, or researcher, email [email protected] to:

• advise on bespoke benchmarks
• provide clean public data, or
• integrate your private data securely.

No cost consultation: https://www.lexpipe.com/meetings/lexpipe/meet-us

---

Time Entry Dataset at a Glance

• 31 litigation matters
• ≈13k unique time entries
• ≈20k hours of billed time
• 4 phases labeled: Pleading, Discovery, Pretrial, Trial

Law firm invoices were OCRed with tersseract 3.0, LLMs extracted time entries, with manual data cleaning. Source PDF documents available on request.

Sample: New York Commercial Contract Case

Time entries for an example matter are shown for a New York commercial contract dispute:

Title	Cowen and Company, LLC v. ReShape Lifesciences, Inc.
Docket	654817/2021
Court	New York State, New York County, Supreme Court
Link	NYSCEF Link to Docket
Summary	Cowen and Company, LLC alleges that ReShape Lifesciences, Inc., breached a 2019 financial advisory agreement by failing to pay $1.35 million in fees after a merger.
Phases	The case advanced through pleading from August 2021, discovery until June 2022, and pretrial motions leading to a money judgment on May 11, 2023, after which it was disposed.

Time entries for the case can be seen below, with the time coded according to the phase.

Sample Rows

---

Detailed Usage

This repository contains both raw and cleaned data.

• Cleaned data CSV file direct download link
• Detailed walkthrough in data analysis notebook
• Raw data for experienced users accessible via Huggingface:

from datasets import load_dataset
repo_id = "lexpipe/time-entries-and-phases"
dataset = load_dataset(repo_id)

Each time entry has an associated duration, and (when available) a name for the phase of legal proceedings. The associated duration is the amount of time logged by the person filing the time entry.

Narratives

Each time entry has a "narrative", a textual descriptions of the work that was conducted by the timekeeper making the entry. Narratives vary in length and detail. In the cleaned data, the narratives have been processed with named entity recognition, with names of people and organizations scrubbed and replaced with "PERSON". The real names of the timekeepers are kept in the respective column.

Certain phrases occur often in the narratives, and can be indicative of the phase of the legal matter. Word size indicates frequency of occurrence in the narratives:

Data Columns

For a description of the columns in the processed data, please see the exploratory data analysis notebook

The (raw) dataset comprises the following columns:

Column	Column Name	Description	Data Type
1	EntryID	Unique identifier of the entry.	Integer
2	Narrative	Description of the work performed during the time entry.	Object (string)
3	PhaseName	Name of the phase in the legal proceedings.	Object (string)
4	PhaseDurationDays	Duration of the legal phase in days.	Float64
5	EntryDurationHours	Total duration of the legal services for the entry (in hours).	Float64
6	Timekeeper	Name of the person logging the time entry.	Object (string)
7	Rate	Hourly rate of the timekeeper.	Float64
8	Position	Position or role of the timekeeper (e.g., associate, partner).	Object (string)
9	Partner	Is the time keeper a partner in the firm or not	Bool
10	Firm	Law firm name associated with the time entry.	Object (string)
11	Date	Date of the time entry.	Object (string)
12	Docket	Unique identifier for the case in the docket system.	Object (string)
13	SourceDocumentZip	Name of the .zip file containing the source document.	Object (string)
14	SourceDocumentName	Name of the source document (PDF).	Object (string)
15	QuerySlug	Unique identifier or slug for the data entry.	Object (string)
16	Length	Length of the document in tokens.	Integer
17	Pages	Pages containing the time entry data.	Object (string)
18	Court	Court presiding over the matter.	Object (string)
19	DocketEntry	Text of the specific docket entry.	Object (string)
20	Link	Link to the docket entry or relevant document.	Object (string)
21	LinkEx	External link related to the time entry.	Object (string)
22	TextMatter	Full or cleaned text related to the matter.	Object (string)
23	DaysFromStart	Days from the start of the case to the time entry.	Float64
24	Matter	Identifier or name of the legal matter.	Object (string)
25	Cleaned	Has the entry been cleaned or not.	Bool
26	Confidence	Confidence score for the extracted data.	Object (string)
27	Analysis	Additional analysis or notes on the time entry.	Object (string)

Statistics of data columns

Column Name	Mean	Median	Range (max minus min)	Cardinality
PhaseDurationDays	365.15	367.00	2023.00	31
EntryDurationHours	1.49	0.70	17.60	172
Rate	578.78	590.00	1507.00	180
Length	268666.61	201498.00	571675.00	46
DaysFromStart	449.63	435.00	1942.00	613

Data Splits

No predefined splits for machine learning purposes have been made, users may create splits (e.g., train/test/validation) as needed.

If treated cross-sectionally, predictive perforamnce of time entry narrative to phase name label drops progressively. From roughly ~2000 narratives predictive performance is adequate.

Contributions

• Michael Sander, CEO / Founder, LexPipe, Inc.

• Ruud van den Brink, Principal Data Scientist, LexPipe, Inc.

Additional Information

Released under Apache-2.0, which gives broad commercial & research rights. Please attribute “LexPipe” and link back to this repo.

https://lexpipe.com