NYC Rental Lease Clause Disguised Fee Anomaly Detector
In response to landlord fee restrictions, non-standard fee clauses have proliferated in residential leases. Build a natural language processing anomaly detector using HuggingFace Transformers to scan lease agreement text and highlight illegal tenant fee surcharges. Achieve 92%+ F1-score on legal phrase classification.
What you are building
Develop an NLP lease agreement auditor using HuggingFace Transformers to detect disguised broker fee clauses.
How work is evaluated
Evaluates sentence-level clause anomaly detection accuracy on standard lease agreements.
Datasets
Public datasets, plus the private datasets you have access to.
How submissions are scored
Flag Prohibited Clause
Must flag non-refundable facilitation fee as prohibited
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
Clause Classification F1-Score
F1 score across test dataset of lease clauses • target: 0.93 • range: 0.85-1
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
Learning goals
Fine-tune a DeBERTa or RoBERTa classification model on lease addendums
Detect suspicious wording like 'mandatory processing surcharge' or 'application review fee'
Extract dollar amounts attached to questionable contract clauses
Output risk-ranked audit summaries for tenant advocacy applications
Resources
Residential lease addendums and fee schedule text blocks collected across NYC borough rental listings.
Challenge setup
Evaluates sentence-level clause anomaly detection accuracy on standard lease agreements.
Challenge input
JSON with clause_text
HuggingFace Transformers
Provides pre-trained models for sequence classification and NER
Evaluated output
JSON with is_prohibited (boolean), anomaly_category, fee_amount
- Must flag non-refundable facilitation fee as prohibited
- F1 score across test dataset of lease clauses • target: 0.93 • range: 0.85-1
- Benchmark: LegalNLPBench
- Clause Classification F1 Score target: 0.93
- 1 public reference case
- Python execution harness
- Python sandbox (unavailable on Versalist)
Shows the first task, up to three tools, and up to three evidence items.
View technical recipe
Configured tools
- Transformers · Required
- Hugging Face · Optional
- Kubeflow · Optional
Evaluation contract
- Flag Prohibited Clause · Weight 1
- Clause Classification F1-Score · Weight 1
Metrics
- Clause Classification F1-Score · Target 0.93
Benchmarks
- LegalNLPBench
- general
Reference cases
- 1 public reference case
Workflow tasks
- Task 1
Input: JSON with clause_text
Output: JSON with is_prohibited (boolean), anomaly_category, fee_amount
Execution harness: Python
Evaluation tools
- HuggingFace Transformers · NLP Framework
Recipe freeze marker
No recipe freeze time is recorded.