LegalBERT, a BERT-based transformer trained on 12 GB of legal text, reaches a macro-F1 of 0.70 on statutory-interpretation classification, outperforming GPT-4’s 0.45 on the same task. The broader legal AI software market hit $4.59 billion in 2025 and is on track to reach $5.59 billion in 2026, according to ResearchAndMarkets. This article covers LegalBERT’s benchmark scores, model specifications, how it compares to general-purpose language models, and the adoption data shaping AI ethics in the legal profession.
LegalBERT Statistics in 2026 – TL;DR
LegalBERT has 110 million parameters and was pre-trained on roughly 355,000 legal documents from six jurisdictions.
On the LexGLUE benchmark, Legal-BERT and CaseLaw-BERT outperform generic transformers by 2 to 4 percentage points in macro-F1 on U.S.-focused tasks.
Fine-tuned LegalBERT scores 95.08% accuracy on legal opinion classification, compared to 94.15% for base BERT.
A hybrid LegalBERT model achieved a 93.4% F1 score on legal entity recognition across 15,000 annotated documents.
The legal AI software market grew from $4.59 billion in 2025 to a projected $5.59 billion in 2026, a 22.3% year-over-year increase based on ResearchAndMarkets data.
What Is LegalBERT?
LegalBERT is a family of BERT-based models pre-trained on English-language legal corpora. The original LEGAL-BERT was developed by Chalkidis et al. (2020) at the Athens University of Economics and Business. It adapts the standard BERT architecture, 12 hidden layers, 768 hidden dimensions, 12 attention heads, for legal language.
The model was trained using Masked Language Modeling (MLM) and Next Sentence Prediction (NSP), the same objectives as base BERT, but applied to legal text. A separate variant, CaseLaw-BERT, was initialized from BERT-base and further trained on 37 GB of U.S. case law from the Harvard Law case corpus, covering 3.4 million decisions from 1965 onward.
LegalBERT Model Specifications
| Parameter | LEGAL-BERT-SC | CaseLaw-BERT |
|---|---|---|
| Architecture | BERT-base | BERT-base |
| Parameters | ~110M | ~110M |
| Hidden Layers | 12 | 12 |
| Hidden Size | 768 | 768 |
| Attention Heads | 12 | 12 |
| Vocab Size | 30,000 | 30,522 |
| Training Data | 12 GB (multi-jurisdiction) | 37 GB (U.S. case law) |
| Pre-training Strategy | From scratch | Further pre-training |
Source: Chalkidis et al. (2020); Zheng et al. (2021)
LEGAL-BERT-SC was trained from scratch on a custom 30,000-subword vocabulary built from legal text. CaseLaw-BERT reused the original BERT vocabulary and ran an additional 1 million pre-training steps on legal data. Both approaches target the same downstream tasks but differ in how they handle domain-specific model training.
LegalBERT Training Corpus Breakdown
LEGAL-BERT-SC draws on six major sources of English legal text. The largest single component is U.S. contracts filed with the SEC via EDGAR, at 3.9 GB. U.S. court cases contribute 3.2 GB. The remaining 4.4 GB comes from EU legislation, UK legislation, European Court of Justice cases, and European Court of Human Rights cases.
| Source | Documents | Size (GB) |
|---|---|---|
| U.S. Contracts (SEC-EDGAR) | 76,366 | 3.9 |
| U.S. Court Cases | 164,141 | 3.2 |
| EU Legislation (EUR-Lex) | 61,826 | 1.9 |
| UK Legislation | 19,867 | 1.4 |
| ECJ Cases | 19,867 | 0.6 |
| ECHR Cases | 12,554 | 0.5 |
Source: Chalkidis et al. (2020)
LegalBERT Training Corpus by Size
How Does LegalBERT Perform on Legal NLP Benchmarks?
LegalBERT’s main evaluation framework is LexGLUE, a benchmark of seven legal NLP tasks spanning human rights law, U.S. case law, EU regulation, and contract law. On tasks tied to U.S. case law, such as SCOTUS classification and CaseHOLD, legal-domain models gained 2 to 4 points in macro-F1 over generic transformers like BERT-base and RoBERTa.
On the statutory-interpretation classification task (a three-way label problem using U.S. Supreme Court opinions), LEGAL-BERT-SC reached a macro-F1 of 0.70. BERT-base scored 0.68. DistilBERT scored 0.67. GPT-4 with a codebook prompt scored just 0.22, and few-shot GPT-4 hit 0.45, based on Thalken et al. (2023). These gaps are significant because they show fine-tuned smaller models can still beat much larger general-purpose models on highly specialized classification tasks.
LegalBERT Statistics: Benchmark Scores Compared
| Model | Task | Metric | Score |
|---|---|---|---|
| LEGAL-BERT-SC | Statutory Interpretation (3-way) | Macro-F1 | 0.70 |
| BERT-base | Statutory Interpretation (3-way) | Macro-F1 | 0.68 |
| DistilBERT | Statutory Interpretation (3-way) | Macro-F1 | 0.67 |
| GPT-4 (few-shot) | Statutory Interpretation (3-way) | Macro-F1 | 0.45 |
| GPT-4 (codebook) | Statutory Interpretation (3-way) | Macro-F1 | 0.22 |
| Legal-BERT | Legal Opinion Classification | Accuracy | 95.08% |
| Legal-Vocab-BERT | Legal Opinion Classification | Accuracy | 95.28% |
| BERT-base | Legal Opinion Classification | Accuracy | 94.15% |
| LegalBERT (ECtHR-B) | Human Rights Violation | Macro-F1 | 74.7% |
| Hybrid LegalBERT | Legal Entity Recognition | F1 | 93.4% |
Source: Thalken et al. (2023); Khan (2021); Chalkidis et al. (2022)
Statutory Interpretation: Macro-F1 by Model
LegalBERT vs. Large Language Models
On LexGLUE tasks, zero-shot GPT-3.5 underperformed fine-tuned Legal-BERT on every task in the benchmark, according to Chalkidis (2023). Prompt-only LLMs like FLAN-T5 and Llama-2-chat scored even lower, around 0.19 to 0.20 on statutory-interpretation classification where LEGAL-BERT-SC hit 0.70.
A 2025 study found that an 8-billion-parameter LLM lightly pre-trained on legal tasks outperformed GPT-4 on some legal tasks, as reported by Dominguez-Olmedo et al. The pattern holds across legal NLP: domain-specific fine-tuning on a small model often beats prompting a much larger one. LLaMA-UTP achieved 90.5% accuracy classifying unfair terms, against LegalBERT’s 82.0% and GPT-3.5’s 78.4%, according to a recent ResearchGate study. This suggests the gap narrows with newer, larger models that also receive domain fine-tuning through cloud AI services.
Unfair Terms Classification Accuracy
LegalBERT Statistics on Downstream Legal Tasks
LEGAL-BERT-SC improved micro-F1 on EURLEX57K multi-label classification by 0.2% over BERT-base. On ECHR binary classification, the gain was 0.8%, and on ECHR multi-label it was 2.5%. For Contracts-NER, gains ranged between 1.1% and 1.8% across entity types. These numbers come from Chalkidis et al. (2020).
A 2025 study on Indian legal judgment summarization compared LEGAL-BERT against BiLSTM. BiLSTM outperformed LEGAL-BERT on ROUGE-1 (45.12 vs. 28.25) and precedent accuracy (93% vs. 78%). The results suggest LEGAL-BERT’s pre-training on Western legal text does not automatically transfer to non-Western jurisdictions, creating demand for region-specific variants like localized AI models in new markets.
How Big Is the Legal AI Market?
The AI-in-legal market reached $4.59 billion in 2025 and is projected at $5.59 billion for 2026, per ResearchAndMarkets. That represents a 22.3% CAGR. By 2030, the same report projects $12.49 billion. Fortune Business Insights puts the legal AI software market at $5.21 billion in 2026, growing to $40.94 billion by 2034 at a 29.4% CAGR.
NLP is the leading technology segment. Future Market Insights estimates NLP holds a 35.7% share of the legal AI technology segment. Machine learning and deep learning together accounted for over 63% of the market in 2024, based on a separate Grand View Research estimate. North America leads regionally with 42% to 46% of global revenue.
| Source | 2025 Value | 2026 Value | 2030 Projection | CAGR |
|---|---|---|---|---|
| ResearchAndMarkets | $4.59B | $5.59B | $12.49B | 22.3% |
| Fortune Business Insights | $4.02B | $5.21B | – | 29.4% |
| MarketsandMarkets | $3.11B | – | $10.82B | 28.3% |
| Future Market Insights | $2.1B | – | – | 13.1% |
Source: ResearchAndMarkets (2026); Fortune Business Insights (2025); MarketsandMarkets (2025)
Legal AI Market Size Growth (ResearchAndMarkets)
Legal AI Adoption Statistics Among Law Firms
Generative AI adoption among legal professionals jumped from 31% in 2025 to 69% in 2026, according to the 8am 2026 Legal Industry Report based on a survey of over 1,300 practitioners. Legal-specific AI tool use doubled to 42%. About 59% of legal professionals now use AI at least several times a week.
Firm-level adoption lags behind individual use. Only 34% of firms have formally adopted AI, and 43% have no AI policy and no plans to create one. Large firms with 51 or more attorneys show a 39% generative AI adoption rate, roughly double that of smaller firms. Law firm technology spending grew 9.7% in 2025, the fastest rate recorded, per the Thomson Reuters/Georgetown Law report. These trends are connected to broader shifts in AI liability and risk management.
| Metric | 2024 | 2025 | 2026 |
|---|---|---|---|
| Personal GenAI Use (Legal Professionals) | 27% | 31% | 69% |
| Legal-Specific AI Tool Use | – | ~21% | 42% |
| Firm-Level AI Adoption | – | ~21% | 34% |
| In-House Counsel AI Use | 23% | – | 52% |
Source: 8am 2026 Legal Industry Report; ACC/Everlaw
GenAI Adoption Among Legal Professionals
LegalBERT Statistics: Use Cases in Legal NLP
The top AI use cases in law are document review (77%), legal research (74%), document summarization (74%), and brief drafting (59%), based on 2025 survey data. LegalBERT and its variants are used in several of these areas, particularly for classification and entity extraction tasks where fine-tuned transformers perform well.
Contract clause classification is one area where fine-tuned Legal-BERT paired with GPT-4o has been tested. The workflow uses LegalBERT to classify clauses by type, then passes them to GPT-4o for risk analysis. Fine-tuned models showed balanced precision, recall, and F1 across k-fold cross-validation, while the untrained Legal-BERT model struggled with accuracy and precision. Researchers working on AI content generation are applying similar hybrid approaches across multiple industries.
LegalBERT Variants Across Languages
Domain-adapted BERT models now exist for legal text in multiple languages. LegalBert-pt covers Brazilian Portuguese law. AraLegal-BERT targets Arabic legal text. InLegalBERT and CustomInLegalBERT were trained on Indian legal corpora, starting from LEGAL-BERT-SC and running 300,000 additional pre-training steps.
InLegalBERT outperformed both the original LegalBERT and CaseLaw-BERT on three Indian legal tasks: statute identification, rhetorical role segmentation, and judgment prediction. These gains confirm that jurisdiction-specific training data matters more than raw model size.
| Model | Language / Jurisdiction | Base Model |
|---|---|---|
| LEGAL-BERT-SC | English (EU, UK, US) | Trained from scratch |
| CaseLaw-BERT | English (U.S.) | BERT-base + 1M steps |
| InLegalBERT | English (India) | LEGAL-BERT-SC + 300K steps |
| LegalBert-pt | Portuguese (Brazil) | BERT-base + legal corpus |
| AraLegal-BERT | Arabic | BERT-base + legal corpus |
| LegalRoBERTa | English | RoBERTa + legal corpus |
| PoL-BERT-Large | English (U.S.) | BERT-large + 256 GB legal text |
Source: Various; Hugging Face model cards
What Are the Limitations of LegalBERT?
Inter-annotator agreement on legal classification tasks is moderate. Krippendorff’s alpha sits around 0.63 even among domain experts, meaning human annotators disagree on about a third of borderline cases. This ceiling constrains how high any model can score.
LegalBERT’s 512-token input limit is another constraint. Legal documents frequently run thousands of tokens long. Hierarchical variants (HierBERT) address this by processing documents in chunks, but they add complexity. Longformer and BigBird handle longer contexts natively, though they don’t consistently outperform LegalBERT on LexGLUE tasks.
Training cost is a practical concern. LegalBERT and similar models do not always outperform base BERT by large margins. On some tasks, the improvement is under 1%. Researchers have questioned whether the cost of collecting legal corpora and running domain-adaptive pre-training is justified when gains are small, an issue that applies broadly to cloud-based AI model training.
FAQ
How many parameters does LegalBERT have?
LegalBERT has approximately 110 million parameters, matching the BERT-base architecture with 12 hidden layers, 768 hidden dimensions, and 12 attention heads.
Does LegalBERT outperform GPT-4 on legal tasks?
On statutory-interpretation classification, fine-tuned LEGAL-BERT-SC scored a macro-F1 of 0.70 versus GPT-4’s 0.45 in few-shot mode, based on Thalken et al. (2023) benchmarks.
What is the LexGLUE benchmark?
LexGLUE is a collection of seven English legal NLP datasets covering document classification, multi-label classification, and multiple-choice tasks across EU, UK, and U.S. law.
How large is the legal AI market in 2026?
ResearchAndMarkets projects the AI-in-legal market at $5.59 billion in 2026, growing at a 22.3% CAGR. Other estimates range from $2.1 billion to $5.21 billion depending on scope.
Is LegalBERT available for free?
Yes. Multiple LegalBERT variants are available on Hugging Face under open licenses, including nlpaueb/legal-bert-base-uncased and casehold/legalbert.
Sources:
https://arxiv.org/pdf/2010.02559
https://www.emergentmind.com/topics/legal-bert
https://www.llamalab.ai/blog/legal-ai-adoption-doubles-2026-8am-report
