Semaj90/gemma4-e4b-legal-grpo

Gemma 4 E4B Legal GRPO — LoRA Adapter

Fine-tuned Google Gemma 4 E4B for legal analysis using GRPO (Group Relative Policy Optimization) reinforcement learning.

Model Details

Property	Value
Base model	google/gemma-4-E4B-it (4B effective params)
Quantized base	unsloth/gemma-4-E4B-it-unsloth-bnb-4bit
Method	GRPO (RL-based alignment)
LoRA rank	16 (alpha=16, RSLoRA)
Trainable params	42.4M (0.53% of base)
Adapter tensors	588 (language_model only)
Adapter size	140 MB
Training hardware	NVIDIA RTX PRO 6000 Blackwell (95 GB, Colab G4)
Training time	~1.3 hours (250 steps, 1000 prompts)
License	Apache 2.0

Model Description

Legal-domain LoRA adapter trained with GRPO reinforcement learning on 45K+ legal/reasoning examples. Optimized for legal document analysis, statute interpretation, evidence classification, and case law reasoning with proper Bluebook citations.

• Developed by: Semaj90
• Model type: PEFT LoRA adapter (text generation)
• Language: English
• License: Apache 2.0
• Fine-tuned from: google/gemma-4-E4B-it

Model Sources

• Repository: Semaj90/gemma4-e4b-legal-grpo
• GRPO Paper: DeepSeekMath (arXiv:2402.03300)

How to Get Started

With Unsloth (recommended)

from unsloth import FastVisionModel
from peft import PeftModel

model, tokenizer = FastVisionModel.from_pretrained(
    model_name="unsloth/gemma-4-E4B-it-unsloth-bnb-4bit",
    max_seq_length=4096,
    load_in_4bit=True,
)
model = PeftModel.from_pretrained(model, "Semaj90/gemma4-e4b-legal-grpo")
FastVisionModel.for_inference(model)

messages = [{"role": "user", "content": [{"type": "text", "text": "Analyze 42 U.S.C. Section 1983"}]}]
inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to("cuda")
output = model.generate(input_ids=inputs, max_new_tokens=512, temperature=0.7)
print(tokenizer.decode(output[0], skip_special_tokens=True))

With Ollama (GGUF)

# See: Semaj90/gemma4-e4b-legal-grpo-GGUF
ollama create gemma4-legal:latest -f Modelfile
ollama run gemma4-legal:latest "What are the elements of negligence?"

Uses

Direct Use

• Legal document analysis and case law research
• Evidence classification and chain of custody assessment
• Statute interpretation (U.S.C., CFR, state codes)
• Case law reasoning and precedent analysis
• Contract review and liability assessment
• Legal AI chatbot / RAG pipeline augmentation

Out-of-Scope Use

• Not a substitute for professional legal advice
• Not designed for jurisdictions outside U.S. law
• Should not be used for automated legal decision-making without human review

Training Details

Training Data

Dataset	Samples	Purpose
FineTome-100k	10,000	General instruction following
GSM8K	5,000	Mathematical reasoning
Pile of Law	20,000	Legal text corpus
LexGLUE CaseHold	5,000	Legal reasoning / holdings
LexGLUE SCOTUS	5,000	Supreme Court opinions
Custom codebase patterns	250	Legal AI system patterns

45K+ examples distilled into 1,000 GRPO prompts (Phase 1 pilot).

Training Procedure

Method: GRPO (Group Relative Policy Optimization) — generates multiple completions per prompt, scores them with reward functions, and updates the policy to favor higher-reward outputs.

Training Hyperparameters

• Training regime: BFloat16
• Optimizer: AdamW 8-bit
• Learning rate: 5e-6 (cosine schedule)
• Batch size: 4 (gradient accumulation: 2, effective: 8)
• Generations per prompt: 2
• Max completion length: 256 tokens
• Warmup steps: 30
• Weight decay: 0.01
• Max grad norm: 0.5
• Epochs: 1

Reward Functions

Single combined reward with 5 signals optimized for training throughput:

Signal	Weight	Description
Citation accuracy	0.25	Valid Bluebook, U.S.C., CFR citation formats
Reasoning logic	0.25	Logical connectors (therefore, because, pursuant to)
Legal formatting	0.20	Numbered lists, paragraph structure
Anti-hallucination	0.15	Penalize fabricated case names
Length efficiency	0.15	Target 100-250 words

Speeds, Sizes, Times

• Total training time: ~1.3 hours (4,744 seconds)
• Steps: 250
• Throughput: 0.21 samples/sec
• Hardware: NVIDIA RTX PRO 6000 Blackwell Edition (95 GB VRAM)
• Cloud: Google Colab G4 GPU

Bias, Risks, and Limitations

• May hallucinate case citations despite anti-hallucination reward signal
• Trained primarily on U.S. federal law — limited state and international coverage
• Text-only adapter — vision/audio capabilities of base Gemma 4 pass through unmodified
• Phase 1 pilot training (1K prompts, 1 epoch) — production training with full dataset recommended

Adapter Surgery Note

The original training produced 884 tensors (588 language + 224 vision + 72 audio). Despite finetune_vision_layers=False, generic target_modules (q_proj, k_proj, etc.) matched projections across all sub-models. The vision/audio tensors use Gemma4ClippableLinear which PEFT cannot merge (upstream issue).

This adapter has been surgically cleaned to contain only the 588 language_model tensors (140 MB). Vision and audio capabilities pass through from the base model unmodified.

Environmental Impact

• Hardware Type: NVIDIA RTX PRO 6000 Blackwell
• Hours used: ~1.3
• Cloud Provider: Google Colab
• Compute Region: US

Technical Specifications

Compute Infrastructure

• Hardware: NVIDIA RTX PRO 6000 Blackwell Edition (95 GB, Compute 12.0)
• Software: Unsloth 2026.4.2, TRL 0.24.0, PEFT 0.18.1, Transformers 5.5.0, PyTorch 2.10.0+cu128

Citation

@misc{gemma4-legal-grpo-2026,
  title={Gemma 4 E4B Legal GRPO Adapter},
  author={Semaj90},
  year={2026},
  url={https://huggingface.co/Semaj90/gemma4-e4b-legal-grpo}
}

Framework Versions

• PEFT 0.18.1
• Unsloth 2026.4.2
• TRL 0.24.0
• Transformers 5.5.0
• PyTorch 2.10.0+cu128