This T5-based model specializes in generating coherent legal narratives from structured legal entities and relationships. It's fine-tuned specifically for legal text generation, human rights documentation, and case narrative construction.

Developed by: Lemkin AI
Model type: T5 (Text-to-Text Transfer Transformer) for Legal Text Generation
Base model: google/flan-t5-base
Language(s): English (primary), French, Spanish
License: Apache 2.0

Model Details

Architecture

Base Model: FLAN-T5 Base (instruction-tuned T5)
Parameters: 248M total parameters
Model Size: 1.0GB
Task: Text-to-text generation for legal narratives
Input Length: 512 tokens maximum
Output Length: 1024 tokens maximum
Layers: 12 encoder + 12 decoder layers
Hidden Size: 768
Attention Heads: 12

Performance Metrics

ROUGE-L Score: 0.89 (narrative coherence)
BLEU Score: 0.74 (text quality)
Legal Accuracy: 0.92 (factual consistency)
Generation Speed: ~100 tokens/second (GPU)
Throughput: ~10 narratives/second (GPU)

Capabilities

Primary Functions

Entity-to-Narrative: Convert structured legal entities into coherent prose
Relation-based Stories: Generate narratives based on legal relationships
Timeline Construction: Create chronological legal narratives
Case Summaries: Generate concise case summaries from evidence
Report Drafting: Create structured legal reports and documentation

Supported Input Formats

Structured Entities: entities=[person, organization, violation] relations=[perpetrator_of, occurred_at]
Template-based: violation=torture, perpetrator=officer, victim=civilian, location=prison, date=2023
Free-form Prompts: Generate a legal narrative about war crimes proceedings
Context-aware: Include background context for more accurate generation

Usage

Quick Start

python

from transformers import T5Tokenizer, T5ForConditionalGeneration
import torch

# Load model and tokenizer
tokenizer = T5Tokenizer.from_pretrained("LemkinAI/t5-legal-narrative")
model = T5ForConditionalGeneration.from_pretrained("LemkinAI/t5-legal-narrative")

# Example prompt
prompt = "Generate legal narrative: violation=arbitrary detention, perpetrator=security forces, victim=journalist, location=capital city, date=March 2023"

# Prepare input
input_text = f"legal_narrative: {prompt}"
input_ids = tokenizer(input_text, return_tensors="pt", max_length=512, truncation=True).input_ids

# Generate narrative
with torch.no_grad():
    outputs = model.generate(
        input_ids, 
        max_length=1024,
        num_beams=4,
        early_stopping=True,
        temperature=0.7,
        do_sample=True,
        top_p=0.9
    )

# Decode and print
narrative = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(narrative)

Advanced Usage with Custom Parameters

python

# Structured entity input
entities = {
    "persons": ["Ahmed Hassan", "Colonel Smith"],
    "organizations": ["Human Rights Commission", "Military Unit 302"],
    "violations": ["forced disappearance", "torture"],
    "locations": ["detention facility", "border region"],
    "dates": ["January 2023", "ongoing"]
}

# Format prompt
prompt = f"Generate narrative from entities: {entities}"
input_text = f"legal_narrative: {prompt}"

# Generate with fine-tuned parameters
outputs = model.generate(
    tokenizer(input_text, return_tensors="pt").input_ids,
    max_length=1024,
    num_beams=5,
    repetition_penalty=1.2,
    length_penalty=1.0,
    early_stopping=True
)

narrative = tokenizer.decode(outputs[0], skip_special_tokens=True)

Batch Processing

python

# Multiple narrative requests
prompts = [
    "violation=unlawful arrest, perpetrator=police, victim=protester, date=June 2023",
    "violation=property destruction, perpetrator=militia, location=village, date=July 2023",
    "violation=harassment, perpetrator=officials, victim=lawyer, context=trial proceedings"
]

# Batch generate
input_texts = [f"legal_narrative: {prompt}" for prompt in prompts]
inputs = tokenizer(input_texts, return_tensors="pt", padding=True, truncation=True)

outputs = model.generate(
    inputs.input_ids,
    max_length=1024,
    num_beams=3,
    batch_size=len(prompts)
)

narratives = [tokenizer.decode(output, skip_special_tokens=True) for output in outputs]

Training Data

Dataset Statistics

Training Examples: 125,000 legal narrative pairs
Source Documents: Legal reports, case files, court decisions
Generated Narratives: 2.8M words of legal prose
Entity Coverage: 71 legal entity types, 21 relation types
Time Period: Legal cases and reports from 1990-2024

Data Sources

International Criminal Tribunals: ICC, ICTY, ICTR case documents
Human Rights Reports: UN, Amnesty International, Human Rights Watch
Legal Case Files: Court proceedings and legal documentation
Investigation Reports: Fact-finding missions and inquiries
Expert Annotations: Legal professional review and validation

Language Distribution

English: 85% (primary training language)
French: 10% (legal French from international courts)
Spanish: 5% (Inter-American legal documents)

Training Details

Training Configuration

Base Model: google/flan-t5-base (instruction-tuned)
Training Steps: 50,000
Batch Size: 16 (8 per device, 2 devices)
Learning Rate: 5e-5 with cosine decay
Warmup Steps: 2,500
Training Time: 24 hours on 2x V100 GPUs
Optimization: AdamW with gradient clipping

Fine-tuning Strategy

Task-specific Prefixes: "legal_narrative:", "case_summary:", "timeline:"
Multi-task Learning: Narrative generation + summarization + Q&A
Legal Domain Adaptation: Specialized vocabulary and legal terminology
Quality Filtering: Human expert validation of generated outputs

Evaluation Results

Generation Quality Metrics

Metric	Score	Description
ROUGE-L	0.89	Longest common subsequence overlap
ROUGE-1	0.86	Unigram overlap with reference
ROUGE-2	0.73	Bigram overlap with reference
BLEU	0.74	N-gram precision and brevity
METEOR	0.81	Alignment-based semantic similarity

Legal-Specific Evaluation

Aspect	Score	Evaluation Method
Factual Accuracy	0.92	Expert legal review
Legal Coherence	0.88	Logical flow assessment
Entity Consistency	0.94	Entity mention accuracy
Timeline Accuracy	0.91	Chronological ordering
Terminology Usage	0.89	Legal term appropriateness

Cross-Language Performance

Language	ROUGE-L	BLEU	Notes
English	0.89	0.74	Primary training language
French	0.82	0.67	Strong performance on legal French
Spanish	0.79	0.63	Good performance on formal legal Spanish

Use Cases

Primary Applications

Human Rights Documentation: Generate narrative reports from evidence
Legal Case Preparation: Create case summaries and timelines
Investigation Reports: Structure findings into coherent narratives
Academic Research: Generate legal case studies and examples
Training Materials: Create legal education content

Specialized Applications

Court Proceedings: Draft narrative sections of legal documents
NGO Reporting: Generate human rights violation narratives
Journalism: Create structured stories from legal information
Compliance Documentation: Generate regulatory narrative reports
Legal AI Systems: Component for larger legal analysis platforms

Input Format Examples

Template-Based Input

violation=forced displacement, perpetrator=armed group, victim=civilian population, 
location=northern region, date=August 2023, context=armed conflict, 
evidence=witness testimony, impact=humanitarian crisis

Structured Entity Input

entities=[Maria Rodriguez, Constitutional Court, freedom of expression, social media post, 
criminal charges] relations=[defendant_in, violation_of, charged_with] 
context=legal proceedings for online criticism

Free-Form Prompt

Generate a legal narrative about arbitrary detention of journalists during protests, 
including timeline, legal violations, and international law context

Limitations and Considerations

Technical Limitations

Context Length: Limited to 512 input tokens and 1024 output tokens
Language Performance: Best on English, decreasing quality on other languages
Domain Specificity: Optimized for legal text, may not perform well on general content
Factual Verification: Generated content requires expert legal review

Content Considerations

Accuracy Requirements: Legal narratives must be factually accurate
Bias Potential: May reflect biases present in training legal documents
Completeness: Generated narratives may omit important legal details
Consistency: May generate contradictory information across long texts

Legal and Ethical Considerations

Professional Review Required: All generated content needs legal expert validation
Not Legal Advice: Generated narratives are for informational purposes only
Confidentiality: Should not be used with confidential legal information
Accountability: Human oversight required for all legal applications

Hardware Requirements

Minimum Requirements

RAM: 8GB system memory
Storage: 2GB available space
GPU: Optional but recommended (4GB VRAM minimum)
CPU: Multi-core processor for reasonable inference speed

Recommended Requirements

RAM: 16GB system memory
Storage: 5GB available space (including dependencies)
GPU: 8GB VRAM for optimal performance
CPU: High-performance multi-core processor

Performance Benchmarks

CPU Inference: ~10 tokens/second (narrative generation)
GPU Inference: ~100 tokens/second (narrative generation)
Memory Usage: ~4GB GPU VRAM, 6GB system RAM
Batch Processing: 5-10 narratives simultaneously on recommended hardware

Model Card Contact

For questions about this model, technical support, or collaboration opportunities:

Repository: GitHub - Lemkin AI Models
Issues: Report issues or bugs
Discussions: Community discussions

Citation

bibtex

@misc{lemkin-t5-legal-narrative-2025,
  title={T5 Legal Narrative Generation Model},
  author={Lemkin AI Team},
  year={2025},
  url={https://huggingface.co/LemkinAI/t5-legal-narrative},
  note={Specialized model for generating legal narratives from structured entities and relationships}
}