Vrishab80/apex-document-relevance-review

Apex Telecommunications Document Relevance Review (Synthetic)

A fully synthetic eDiscovery / internal-investigation dataset built to test whether an AI agent can perform first-pass relevance coding on a realistic, adversarial document population.

The scenario

Apex Telecommunications, Inc. suspects Senior Sales Manager Michael Carter of sharing confidential pricing and customer information with a competitor, Northstar Communications, before and during his departure from the company. The dataset contains 138 documents (emails, chat logs, meeting notes, and attachments) spanning the six-month investigation period (Jan 1 - Jun 30, 2025).

All people, companies, and events are fictional.

Why it's adversarial

The document set is deliberately constructed to defeat naive keyword-matching approaches:

• Keyword traps — the competitor's name appears in clearly irrelevant documents, and a similarly-named but unrelated company ("Northgate" vs. "Northstar") is designed to trigger false positives.
• Authorship bias — several benign documents are authored by the subject of the investigation.
• Document families — one relevant document's evidence lives entirely in an attachment referenced by a generic "see attached" email.
• Buried evidence — a genuinely relevant admission is nested inside a multi-level forwarded email chain.
• Genuine ambiguity — a set of documents are honestly unclear even to a human reviewer, testing whether a reviewer (human or AI) has the judgment to escalate rather than force a confident guess.

Labels

Each document is labeled with one of three classes:

Label	Count	Meaning
Relevant	30	Contains evidence of disclosure, competitor communication, intent, or concealment
Not Relevant	80	Routine, unrelated business communication
Needs Human Review	28	Genuinely ambiguous — should be escalated, not guessed

Files

• data.jsonl — one row per document: doc_id, text, date, author, recipient, cc, type, subject, label, reason. This is the file the Hugging Face Dataset Viewer will render and what load_dataset() picks up automatically.
• documents/ — the original raw documents as individual .txt files, in the format an agent would actually be handed (headers + body), for use in agentic/tool-use evaluations rather than plain classification.
• document_metadata.csv — the metadata table on its own.
• ground_truth.csv — the label/reason table on its own.
• review_protocol.docx — the relevance criteria reviewers (human or AI) were asked to apply.
• score_output.py — scores a model's predictions (ai_output.csv) against ground_truth.csv, reporting accuracy, precision, recall, and F1, with Needs Human Review treated as a correct escalation rather than a miss.

Loading

from datasets import load_dataset

ds = load_dataset("Vrishab80/apex-document-relevance-review")
print(ds["train"][0])

Intended use

Built as a course project testing whether a frontier AI agent can perform accurate, well-reasoned first-pass document relevance review. Useful as a small benchmark for agentic document review, legal-tech / eDiscovery tooling evaluation, or prompt-engineering experiments around grounded classification with an abstain option.