AmirMohseni/WildChat-Legal-Classification-V2-Balanced

WildChat Legal Classification (Balanced)

English multi-turn conversations from WildChat-1M, silver-labeled with GPT-5.4 for legal-guidance detection and legal topic assignment. Rows are class-balanced: equal counts of seeks_legal_guidance=true and seeks_legal_guidance=false.

Use this dataset to train or evaluate models that must (1) decide whether a user is seeking legal guidance and (2) assign a primary legal topic when they are.

Related datasets

Dataset	Contents
AmirMohseni/WildChat-Legal-Classification-V2-LegalOnly	All legal-guidance rows only, with extracted incidents nested per conversation
AmirMohseni/WildChat-Legal-Questions	One row per extracted standalone legal question (flattened from incidents)

Splits

Each split is available as a config key (total, train, val, test). Partition ratio: 90 / 5 / 5 (shuffle seed 42).

Split	Examples
total	2,122
train	1,909
val	106
test	107

Schema

Column	Type	Notes
id	string	WildChat conversation id
conversation	list of {role, content}	Multi-turn user–assistant thread
seeks_legal_guidance	bool	Primary label — user asks for legal guidance on a real or practical situation
primary_topic	string	Single best topic; empty string when seeks_legal_guidance=false
topics	list[string]	1–3 reasonable topic labels when guidance is sought; empty otherwise
contains_legal_content	bool	Conversation discusses law, rights, courts, regulations, etc.
personalized_or_operational	bool	User asks about a real/practical situation, not exam or translation tasks
uncertainty	string	Labeler self-reported confidence: low, medium, or high
relevant_user_turn_ids	list[int]	Turn indices supporting the classification
relevant_assistant_turn_ids	list[int]	Assistant turns needed to interpret elliptical user messages

Usage

from datasets import load_dataset

ds = load_dataset("AmirMohseni/WildChat-Legal-Classification-V2-Balanced", split="train")
row = ds[0]
print(row["seeks_legal_guidance"], row["primary_topic"])
print(row["conversation"][0]["content"][:200])

Dataset construction

Source

Conversations are drawn from allenai/WildChat-1M. WildChat is released under ODC-BY; this derivative adds GPT-5.4 silver labels and resampled splits under CC-BY-4.0.

Labeling pipeline

All annotations use GPT-5.4 with structured JSON outputs.

Pass 1 (low effort). GPT-5.4 triages each conversation into legal, non-legal, or maybe legal to narrow the candidate pool (~11k conversations).

Pass 2 (high effort). GPT-5.4 re-labels legal and maybe legal rows with the full schema above (contains_legal_content, seeks_legal_guidance, personalized_or_operational, topics, primary_topic, uncertainty, relevant turn ids). A one-sentence rationale is produced during labeling but is not included in the published dataset.

English filtering. Rows are kept only when user-authored text is English-dominant. GPT-5.4 additionally filters out conversations that appeared English on metadata but were predominantly written in another language.

Balanced sampling. All 1,061 seeks_legal_guidance=true rows are included. An equal number of seeks_legal_guidance=false rows is drawn uniformly at random from the English non-guidance pool (seed 43), yielding 2,122 rows at 50/50 class balance.

A 100-item human audit of GPT-5.4 labels on a manually curated 50/50 sample found ~93% overall agreement with human judgment.

Topic taxonomy

Fourteen labels, kept compact enough for training but separable in practice:

Label	Short description
FAMILY	Marriage, divorce, custody, support, adoption, guardianship, domestic violence.
HOUSING	Rent, eviction, deposits, mortgages, foreclosure, habitability, neighbors, housing subsidies.
WORK	Employment contracts, wages, dismissal, discrimination at work, leave, workplace safety.
PUBLIC_BENEFITS	Unemployment, disability, pensions, welfare, eligibility, appeals on benefits.
CRIMINAL_JUSTICE	Police, arrest, charges, fines, prosecution, defense, victims' rights, criminal procedure.
CONSUMER_DEBT	Purchases, warranties, refunds, scams, debt collection, loans, bankruptcy, credit.
CONTRACTS	Civil agreements; breach/interpretation not better captured by work/housing/consumer/business.
IMMIGRATION	Visas, residence permits, asylum, citizenship, deportation, family migration.
BUSINESS	Company formation, shareholders, commercial compliance, B2B disputes, self-employment.
DATA_PRIVACY	Personal data, surveillance, GDPR, data deletion, consent, platform data practices.
INTELLECTUAL_PROPERTY	Copyright, trademark, patent, trade secrets, licensing, infringement.
CIVIL_RIGHTS	Discrimination outside work/housing, free speech, due process, equal treatment.
INTERNATIONAL_CROSS_BORDER	Choice of law, jurisdiction, treaty-based questions, cross-border enforcement.
OTHER	Genuinely legal but not covered above.

Selection rules applied during labeling:

• Use CONTRACTS only when the issue is mainly about a civil agreement and is not better captured by WORK, HOUSING, CONSUMER_DEBT, or BUSINESS.
• Use INTERNATIONAL_CROSS_BORDER only when the cross-border or jurisdictional aspect is central to the answer, not merely incidental.

Limitations

• Labels are silver annotations from GPT-5.4, not professional legal review. Do not use for legal advice or high-stakes routing without human oversight.
• English-focused; jurisdiction is preserved in text when stated but not normalized.
• The 50/50 class balance is not representative of real WildChat traffic, where legal-guidance queries are much rarer.

Citation

If you use this dataset, please cite WildChat and acknowledge GPT-5.4 silver labeling:

@misc{wildchat-legal-classification-balanced,
  title={WildChat Legal Classification (Balanced)},
  author={Mohseni, Amir},
  year={2025},
  publisher={Hugging Face},
  howpublished={\\url{https://huggingface.co/datasets/AmirMohseni/WildChat-Legal-Classification-V2-Balanced}}
}