---
language: 
  - en
tags:
  - t5
  - text2text-generation
  - text-generation-inference
  - bitsandbytes
  - 8-bit
  - legal
  - humanizer
  - ai-bypass
license: apache-2.0
datasets:
  - custom-legal-human-corpus
metrics:
  - perplexity
  - readability
widget:
  - text: "It is important to note that the court found the defendant guilty of negligence due to a failure to uphold the standard of care."
    example_title: "AI Detection Bypass (Legal)"
---

# 🏛️ Amicus Humanizer v1 (8-bit)

**Amicus Humanizer v1** is a specialized `text2text-generation` model designed to rewrite AI-generated text to sound entirely human. Built on the highly efficient **T5** architecture and quantized to **8-bit precision** using `bitsandbytes`, this model is explicitly fine-tuned to bypass AI detectors while strictly preserving domain-specific meaning, legal citations, and professional tone.

Developed by the team at **Dockase**, this open-source release aims to empower researchers, law firms, and legal tech developers to seamlessly convert robotic, predictable AI drafts into authoritative, human-sounding prose.

---

## 🚀 Key Features

- **AI Detection Bypass:** Restructures syntax, injects "burstiness" (sentence variety), and eliminates predictable AI tropes to seamlessly bypass detectors like GPTZero, Turnitin, and Copyleaks.
- **Domain-Specific Preservation:** Unlike generic paraphrasers, Amicus is trained to *never* alter strict legal terminology, case citations, or core factual arguments.
- **8-Bit Quantization:** Shipped in 8-bit precision (`bitsandbytes`), reducing VRAM requirements drastically. It runs lightning-fast on cheap consumer GPUs (like the NVIDIA T4) with zero degradation in output quality.
- **Active Voice Prioritization:** Automatically converts passive, bloated AI phrasing ("It is crucial to remember that...") into sharp, authoritative active voice.

---

## 🛠️ Usage & Inference

Because the model is quantized using `bitsandbytes`, you can load it instantly with Hugging Face `transformers`. 

### Installation
```bash
pip install transformers accelerate bitsandbytes
```

### Python Example
```python
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

model_id = "WhiteRoomProdigy/amicus-humanizer-v1"

# Load the tokenizer and 8-bit model
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSeq2SeqLM.from_pretrained(model_id, load_in_8bit=True, device_map="auto")

# Input text (e.g., rigid AI-generated legal text)
input_text = "Furthermore, it is important to note that the plaintiff failed to establish a breach of contract."

# Tokenize and Generate
inputs = tokenizer(input_text, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_length=512, do_sample=True, temperature=0.7)

# Decode the humanized output
humanized_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(humanized_text)
```

---

## 🧠 Model Architecture
- **Base Model:** T5 (Text-to-Text Transfer Transformer)
- **Task:** Text-to-Text Generation (Paraphrasing / Humanization)
- **Quantization:** 8-bit (`bitsandbytes`)
- **Language:** English
- **License:** Apache 2.0

---

## ⚖️ Intended Use & Limitations

**Intended Use:**
- Rewriting first-draft AI legal memos into finalized, natural-sounding documents.
- Restructuring rigid prose to improve readability and flow.
- Bypassing false-positive AI detection on legitimate professional writing.

**Limitations:**
- While it strongly preserves meaning, outputs should always be reviewed by a human expert before use in official legal filings.
- The model is primarily trained on English legal and professional text; performance may drop on highly informal slang or non-English languages.

---

## 🏢 About Dockase
This model was developed by **[Dockase](https://dockase.com)**, the fundamental legal operating system and workflow automation stack for the African market. We are committed to open-sourcing powerful, domain-specific AI tools to push the boundaries of legal tech.