Instructions to use Ronin48LLC/atticus-lora-adapter with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Ronin48LLC/atticus-lora-adapter with PEFT:

from peft import PeftModel
from transformers import AutoModelForCausalLM

base_model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.3-70B-Instruct")
model = PeftModel.from_pretrained(base_model, "Ronin48LLC/atticus-lora-adapter")

Transformers

How to use Ronin48LLC/atticus-lora-adapter with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="Ronin48LLC/atticus-lora-adapter")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("Ronin48LLC/atticus-lora-adapter", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Ronin48LLC/atticus-lora-adapter with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Ronin48LLC/atticus-lora-adapter"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Ronin48LLC/atticus-lora-adapter",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Ronin48LLC/atticus-lora-adapter

SGLang

How to use Ronin48LLC/atticus-lora-adapter with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Ronin48LLC/atticus-lora-adapter" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Ronin48LLC/atticus-lora-adapter",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Ronin48LLC/atticus-lora-adapter" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Ronin48LLC/atticus-lora-adapter",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Ronin48LLC/atticus-lora-adapter with Docker Model Runner:
```
docker model run hf.co/Ronin48LLC/atticus-lora-adapter
```

ATTICUS — Advocacy, Trial, Testimony, Innocence, Case, Unified Scout

"You never really understand a person until you consider things from his point of view… until you climb into his skin and walk around in it." — Atticus Finch, To Kill a Mockingbird

An open-source LoRA adapter trained for public defenders.

ATTICUS is a QLoRA adapter on top of Meta Llama 3.3 70B Instruct. Given a case file, charge sheet, or fact pattern, it identifies defense strategies, constitutional violations, evidentiary weaknesses, Brady/Giglio obligations, and mitigating factors — in plain language, with cited authority.

The name is deliberate. Atticus Finch was not a man who won every case. He was a man who believed that every defendant deserved someone in their corner who would look at the evidence honestly, challenge the state's case rigorously, and treat the accused as a human being worthy of a real defense. This model carries that obligation.

⚠ About this upload. The adapter weights here come from a three-step pilot run (trainer_state.json: global_step 3, max_steps 3, 3 epochs, batch size 2). Treat them as an early preview of the training recipe, not a finished model. The full run is tracked in the source repository.

One half of a balanced system

ATTICUS is the counterpart of SELMA, the suite's law-enforcement model. Where SELMA identifies what the prosecution can charge, ATTICUS builds the defense. Where SELMA maps evidence to statutes, ATTICUS maps evidence to constitutional protections. A system that only serves prosecution is a system that can cause harm; ATTICUS ensures that every capability SELMA gives law enforcement has a counterpart in the hands of the public defender.

	SELMA	ATTICUS
Purpose	Prosecution-side statute identification	Defense-side strategy and analysis
Users	Patrol officers, detectives, special agents	Public defenders, defense attorneys
Output	Applicable charges and elements	Defense theories, constitutional violations, evidentiary weaknesses
Training data	Criminal statutes, case law, charging documents	Suppression motions, acquittals, Brady/Giglio material, exoneration data

Model Details

Field	Value
Base Model	meta-llama/Llama-3.3-70B-Instruct
Adapter Type	LoRA (QLoRA)
LoRA Rank / Alpha / Dropout	64 / 128 / 0.05
Target Modules	`q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`
Task	Defense strategy, constitutional analysis, evidentiary review
Training Method	QLoRA supervised fine-tuning via `SFTTrainer` (trl)
License	Apache 2.0 (adapter weights)
Maintainer	Ronin48LLC

Capabilities

Given a charge sheet or incident description, ATTICUS can:

Defense Strategy — viable theories (alibi, self-defense, lack of intent, entrapment, duress, consent, mistake of fact) and the evidence that supports each
Constitutional Analysis — 4th, 5th, 6th, and 14th Amendment violations: unlawful search and seizure, Miranda failures, right-to-counsel violations, due process, selective prosecution
Evidentiary Weaknesses — element failures, chain-of-custody gaps, reliability problems with forensic methods or expert witnesses
Brady/Giglio Material — categories of evidence the prosecution may be obligated to disclose, and the Strickler v. Greene materiality standard
Sentencing Mitigation — U.S.S.G. and state mitigating factors, diversion eligibility, plea alternatives, departures and variances
Lesser Included Offenses — what the jury could convict on instead, and how to request the instruction
Cross-Reference — relevant case law, circuit splits, jurisdictional quirks
Wrongful Conviction Patterns — eyewitness misidentification, false confessions, informant testimony, bad forensics, Brady suppression

Jurisdictions: U.S. Code and the Federal Rules of Criminal Procedure as the baseline, with state criminal codes and constitutions layered on top. Priority states: Georgia, California, Texas, New York, Florida.

Constitutional Override

The U.S. Constitution is the supreme law of the land. ATTICUS is trained to treat it that way. Where a charge, a search, an interrogation, or a prosecution implicates a defendant's constitutional rights, it says so plainly:

⚠ CONSTITUTIONAL CONCERN — this charge or evidence may not survive challenge under the [Amendment]. ATTICUS recommends filing a motion to suppress / dismiss before trial.

Protections covered: First (speech and association), Fourth (searches and seizures, Mapp v. Ohio), Fifth (self-incrimination, Miranda, double jeopardy, grand jury), Sixth (counsel, speedy trial, confrontation, jury), Eighth (bail and punishment), Fourteenth (due process, equal protection, selective prosecution).

Intended Use

For licensed attorneys and supervised law students:

Identifying potential Fourth, Fifth, Sixth, and Fourteenth Amendment violations
Analyzing evidentiary weaknesses in the prosecution's case
Surfacing defense-strategy considerations for case review
Constitutional case-law research, subject to verification
Training and scenario-based learning in law schools and public defender offices

ATTICUS outputs must be reviewed and verified by a licensed attorney before use in any filing, motion, or client communication.

Limitations

Read LIMITATIONS.md before deploying in any operational context. In short:

ATTICUS is not a licensed attorney; its outputs are not legal advice and create no attorney-client relationship.
It cannot read the record — it works from the facts you provide.
Constitutional law varies by circuit, state, and judge; outputs are general unless you specify jurisdiction.
It has a training-data cutoff.
It can hallucinate case citations. Every case name, citation, and holding must be verified in a primary source before use.
It does not know your client, and it cannot replace voir dire, negotiation, or courtroom judgment.

Usage

This is a PEFT LoRA adapter. Load it on top of the base model:

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.3-70B-Instruct",
    load_in_4bit=True,
    device_map="auto",
)
model = PeftModel.from_pretrained(base_model, "Ronin48LLC/atticus-lora-adapter")
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.3-70B-Instruct")

Note: Access to the base model requires accepting Meta's license on Hugging Face.

A prompt-only 8B stand-in (Llama 3.1 8B Instruct with ATTICUS's system prompt, no fine-tuning) is published as Ronin48LLC/atticus.

Training

Framework: PyTorch + Transformers + TRL + PEFT, 4-bit quantized base
Adapter config: rank 64, alpha 128, dropout 0.05, all seven attention and MLP projections (see adapter_config.json)
This run: 3 optimizer steps, 3 epochs, per-device batch size 2 (checkpoint-3/trainer_state.json); final/ holds the same weights as the repository root

Training Data Sources

Source	Description	Size	License
U.S. Code Title 18	Federal criminal statutes (USLM XML)	~2,700 sections	Public Domain
Federal Rules of Criminal Procedure	Procedural rights of defendants	Full text	Public Domain
U.S. Sentencing Guidelines	Sentencing ranges and mitigating factors	Full manual	Public Domain
State Criminal Codes	GA, CA, TX, NY, FL	~2,500 sections	Fair Use
SCOTUS Criminal Rights Opinions	4th, 5th, 6th, 8th, 14th Amendment decisions	~5K opinions	Public Domain
CourtListener	Federal criminal appeals, suppressions, acquittals	~10K opinions	Open
National Registry of Exonerations	Wrongful conviction data	3,000+ cases	Public Domain
Innocence Project Case Summaries	DNA exoneration summaries	375+ cases	Fair Use
LegalBench	Legal reasoning benchmark tasks	91.8K examples	Open
CaseHOLD	Legal holding classification	585K examples	Open
Synthetic	Generated charge-to-defense mappings	~50K examples	Apache 2.0

Related Models — the Ronin 48 suite

Model	Domain	Use when…
ATTICUS (this adapter)	Public defense	Defense strategy, constitutional violations, evidentiary weaknesses
SELMA	Law enforcement	ATTICUS's counterpart — statute identification and charge elements
ABBY	Forensic investigation	Digital and physical evidence, chain of custody, admissibility
BONES	EMS	Patient assessment, protocols, drug dosing, triage
BRUNO	Fire service	Fireground tactics, size-up, hazmat, extrication, ICS

Source

github.com/CryptoJones/ATTICUS — training pipeline, data collection, jurisdiction architecture.

License

Downloads last month: 19

Model tree for Ronin48LLC/atticus-lora-adapter

Base model

meta-llama/Llama-3.1-70B

Finetuned

meta-llama/Llama-3.3-70B-Instruct

Adapter

(443)

this model