You can deploy this model and start earning money today!

israelburns/
jeremy-gemma4

Name: jeremy-gemma4
Availability: InStock

israelburns/jeremy-training-data

Views

No views yet

Jeremy AI — Gemma 4 E4B Legal Procedural Guidance

Kaggle Hackathon

GitHub

Training Notebook

A QLoRA adapter fine-tuned on Gemma 4 E4B for legal procedural guidance. Built for the Gemma 4 Good Hackathon as part of the Pro Se Network — free legal AI infrastructure for self-represented litigants.

Mission

80% of low-income Americans cannot afford a lawyer. Jeremy AI is a nonprofit infrastructure project that gives self-represented litigants a trained legal procedural guide — free, self-hostable, and deployable anywhere in the world at $0 per conversation.

The choice of Gemma 4 open weights is a mission requirement, not a technical preference. A per-token API bill is not acceptable for legal aid organizations.

What Jeremy Does

Jeremy is a legal procedural guidance engine — not a legal advice chatbot. It:

Understands a user's legal situation in plain English
Extracts structured legal facts (dates, entities, deadlines, risk signals)
Explains procedures, deadlines, and next steps in plain English
Helps draft Answers, EEOC Charges, Motions to Dismiss
Routes high-risk situations (custody, criminal) to attorney referral

It runs on top of a deterministic rule engine covering 12 areas of law across 13 U.S. jurisdictions. The LLM never decides the law — rules are loaded from verified JSON databases.

Training

Dataset

5,196 legal Q&A pairs from 5 free sources — israelburns/jeremy-training-data

Source	Pairs
Synthetic (from rule JSONs)	959
law.stackexchange.com	2,805
SEC EDGAR (employment contracts)	202
CourtListener API (case law)	1,119
Court self-help guides	111

Training Configuration

Parameter	Value
Base model	`unsloth/gemma-4-E4B-it-unsloth-bnb-4bit`
Method	QLoRA
Rank (r)	16
Alpha	32
Dropout	0.0
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Quantization	4-bit (NF4 via Unsloth)
Seq length	1024
Epochs	3
Steps	927
Batch size	1 (grad accum 16, effective batch 16)
Learning rate	2e-4 (cosine decay)
Optimizer	adamw_8bit
GPU	Kaggle T4 (15.6GB, sm_75, fp16)
Training time	5h 18m
Cost	$0

Trainable Parameters

42,401,792 of 8,038,558,240 total (0.53%)

Usage

python

from peft import PeftModel
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

base_model = "google/gemma-4-E4B-it"
adapter = "israelburns/jeremy-gemma4"

tokenizer = AutoTokenizer.from_pretrained(base_model)
model = AutoModelForCausalLM.from_pretrained(
    base_model,
    torch_dtype=torch.float16,
    device_map="auto"
)
model = PeftModel.from_pretrained(model, adapter)

messages = [
    {"role": "system", "content": "You are Jeremy, a legal procedural guidance AI for the Pro Se Network. You help self-represented litigants understand legal procedures. You are NOT a lawyer. Procedural guidance only, never legal advice. Always include a disclaimer."},
    {"role": "user", "content": "I was served with a lawsuit in New York. I have 20 days to respond. What do I do?"}
]

inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)
outputs = model.generate(inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Real-World Impact

McDaniel v. City of New York (Bronx County Supreme Court, 2026) — Section 1983 civil rights case against 6 NYPD officers. Jeremy helped draft the complaint, demand letter, and preservation notices.
Multiple pro se litigants guided through EEOC filings, contract disputes, and eviction defense across New York.

Contributor

Israel "Ace" Burns — Cornell Tech LLM '17 · Co-Founder, Pro Se Network · New York

israelburns/jeremy-gemma4