Symio-ai/legal-drafter-federal

Model Description

Legal Drafter Federal generates federal court documents including complaints, answers, motions, and briefs compliant with the Federal Rules of Civil Procedure (FRCP) and applicable local rules. Trained on PACER filings, federal appellate opinions, and CM/ECF formatting requirements.

Intended Use

• Primary: Generate first drafts of federal court filings in GLACIER Stage 4
• Secondary: Handle multi-jurisdictional cases that move to federal court (diversity, federal question)
• Integration: Called during GLACIER Stage 4 for federal jurisdiction filings

Task Type

text-generation -- Instruction-tuned generation of structured legal documents

Base Model

meta-llama/Llama-3.1-8B-Instruct -- Unified base with FL/MS drafters for consistent pipeline

Training Data

Source	Records	Description
PACER Federal Filings	~500K	Complaints, motions, briefs from US District Courts
Federal Circuit Opinions	~1M	All circuit court opinions (emphasis on 11th, 5th Circuits)
US Code	~60K sections	Full federal statutory text
FRCP + FRE	all rules	Federal Rules of Civil Procedure and Evidence
Local Rules	~200 courts	District-specific local rules
CM/ECF Guidelines	all districts	Electronic filing format requirements

Federal-Specific Training

• Caption format: IN THE UNITED STATES DISTRICT COURT FOR THE [DISTRICT]
• FRCP compliance (Rule 8 pleading, Rule 11 certification, Rule 12 motions)
• Page limits per local rules (varies by district)
• CM/ECF formatting (PDF conversion requirements)
• Federal discovery rules (Rule 26, 30, 33, 34, 36)
• Twombly/Iqbal plausibility standard for federal complaints

Benchmark Criteria (90%+ Target)

Metric	Target	Description
Format Compliance	>= 94%	Pass format-guard (federal mode)
FRCP Accuracy	>= 93%	Correct rule references
Plausibility Standard	>= 90%	Complaints meet Twombly/Iqbal standard
Caption Accuracy	100%	Correct federal court caption
Page Limit Compliance	>= 95%	Respects local rule page limits

GLACIER Pipeline Integration

STAGE 1 (Classify) --> jurisdiction detected as Federal
STAGE 4 (First Draft) --> drafter-federal generates v1 document
STAGE 5 (WDC #2) --> validated by format-guard (federal mode) + local rule check

Training Configuration

• LoRA fine-tuning (r=16, alpha=32)
• Epochs: 3
• Learning rate: 1e-4
• Max sequence length: 8192 (federal briefs tend to be longer)
• Hardware: AWS SageMaker ml.g5.12xlarge

Limitations

• Local rules vary significantly across 94 federal districts; not all are fully represented
• Specialized federal courts (bankruptcy, tax, claims) have limited training data
• Does not generate appellate briefs (different format and structure)
• CM/ECF PDF conversion is a post-processing step, not part of model output

Version History

Version	Date	Notes
v0.1	2026-04-10	Initial model card, repo created