Instructions to use changcheng967/Aegis-Qwen3-8B-Humanizer-v1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use changcheng967/Aegis-Qwen3-8B-Humanizer-v1 with Transformers:

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

pipe = pipeline("text-generation", model="changcheng967/Aegis-Qwen3-8B-Humanizer-v1")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("changcheng967/Aegis-Qwen3-8B-Humanizer-v1")
model = AutoModelForCausalLM.from_pretrained("changcheng967/Aegis-Qwen3-8B-Humanizer-v1")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use changcheng967/Aegis-Qwen3-8B-Humanizer-v1 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "changcheng967/Aegis-Qwen3-8B-Humanizer-v1"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "changcheng967/Aegis-Qwen3-8B-Humanizer-v1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/changcheng967/Aegis-Qwen3-8B-Humanizer-v1

SGLang

How to use changcheng967/Aegis-Qwen3-8B-Humanizer-v1 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 "changcheng967/Aegis-Qwen3-8B-Humanizer-v1" \
    --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": "changcheng967/Aegis-Qwen3-8B-Humanizer-v1",
		"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 "changcheng967/Aegis-Qwen3-8B-Humanizer-v1" \
        --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": "changcheng967/Aegis-Qwen3-8B-Humanizer-v1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use changcheng967/Aegis-Qwen3-8B-Humanizer-v1 with Docker Model Runner:
```
docker model run hf.co/changcheng967/Aegis-Qwen3-8B-Humanizer-v1
```

You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Aegis-Qwen3-8B-Humanizer-v1

A DPO-fine-tuned Qwen3-8B model that rewrites AI-generated text to sound naturally human. Part of Project Aegis — an open-source academic defense tool built to protect students from false AI-detection accusations.

What It Does

Given AI-generated text, the model rewrites it to sound like a real student wrote it — natural phrasing, varied sentence structure, conversational but academic tone.

Model Details

Base Model

Reference: Qwen/Qwen3-8B (Apache 2.0)
Architecture: Qwen3-8B (8.2B params, 36 layers, 4096 hidden dim)
Attention: GQA (32 Q heads / 8 KV heads, head_dim=128)
Context Window: 32,768 tokens native
QK-Norm: Built-in RMSNorm on Q/K projections — prevents extreme logits on non-NVIDIA hardware
tie_word_embeddings: false (separate lm_head)

Fine-Tuning

Method: DPO (Direct Preference Optimization) via TRL DPOTrainer
PEFT: LoRA (rank=64, alpha=128, dropout=0.05)
Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Trainable Params: 174,587,904 (2.09% of 8,365,323,264 total)
LoRA Adapter: Merged into base weights (full model, no separate adapter needed)

Training Configuration

Parameter	Value
Hardware	MACA C500 (64GB HBM2e, 280 TFLOPS BF16)
Software	mcPytorch 2.8.0 (PyTorch 2.8 + MACA SDK 3.3)
Precision	BF16
Attention	SDPA
Dataset	732 DPO preference pairs (changcheng967/aegis_rewriter_dataset)
Epochs	3
Batch Size	2 per device × 8 grad accum = 16 effective
Learning Rate	5e-7 (cosine decay, 10% warmup)
Optimizer	AdamW (torch)
Beta (DPO)	0.1
Max Length	3072 tokens
Gradient Checkpointing	Enabled
Max Grad Norm	1.0
Training Time	~1h 39min (138 steps)
Final Loss	0.069
Final Reward Margin	+2.965

Training Curves

Step	Epoch	Loss	Reward Margin	Accuracy
10	0.22	0.706	-0.018	38%
20	0.44	0.654	+0.088	71%
40	0.87	0.442	+0.602	100%
60	1.31	0.238	+1.348	100%
80	1.74	0.128	+2.069	100%
110	2.39	0.071	+2.741	100%
138	3.00	0.069	+2.965	100%

Quick Start

Installation

pip install transformers torch

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "changcheng967/Aegis-Qwen3-8B-Humanizer-v1",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(
    "changcheng967/Aegis-Qwen3-8B-Humanizer-v1",
    trust_remote_code=True,
)

SYSTEM_PROMPT = (
    "You are a text humanizer. Rewrite the following AI-generated text "
    "to sound naturally human while preserving the original meaning. "
    "Use natural phrasing, varied sentence structure, and conversational tone. /no_think"
)

ai_text = """In recent years, artificial intelligence has revolutionized the way we
approach complex problems across various domains. Machine learning algorithms have
demonstrated remarkable capabilities in pattern recognition, natural language
processing, and decision-making tasks."""

messages = [
    {"role": "system", "content": SYSTEM_PROMPT},
    {"role": "user", "content": ai_text + " /no_think"},
]

input_text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(input_text, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=1024,
        temperature=0.7,
        do_sample=True,
        top_p=0.9,
        repetition_penalty=1.1,
    )

response = tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)

# Strip Qwen3 thinking output if present
if "</think>" in response:
    response = response.split("</think>", 1)[-1].strip()

print(response)

Important: Disable Thinking

Qwen3-8B has a built-in chain-of-thought (thinking) mode. To get clean humanized output without the reasoning step:

Add /no_think to the system prompt and user message
Strip any remaining ({...}) tags from the output (see code above)

Limitations

v1 is DPO-only — the model learned to prefer human text over AI text, but its generation style still carries AI-like patterns. A future SFT + DPO two-stage release will produce more convincing humanized output.
Best results on essay-length English text (500–1500 words).
Not tested on non-English text.
The 732-sample training dataset covers a limited range of topics. Broader data will improve generalization.

Intended Use

Built for students who:

Use AI as a brainstorming or tutoring tool but write their own final drafts
Get falsely flagged by AI detectors (GPTZero, Turnitin, etc.) despite writing original work
Want to verify their natural writing style won't trigger false positives

Not intended for submitting AI-generated work as entirely your own. Aegis is a shield, not a cheat code.

Dataset Format

The model was trained on DPO preference pairs in JSONL format:

{
  "prompt": "Rewrite this text to sound more human: [AI-generated text]",
  "chosen": "[Human-written student essay]",
  "rejected": "[AI-generated essay]"
}

Related Projects

Qwen3-8B — Base model by Qwen Team (Apache 2.0)
TRL — Transformer Reinforcement Learning library (DPOTrainer)
PEFT — Parameter-Efficient Fine-Tuning (LoRA)

Roadmap

SFT training on human-written student essays
SFT + DPO two-stage pipeline (v2)
Broader training dataset (more topics, writing styles, academic levels)
Quantized versions (GGUF, GPTQ) for local inference on consumer hardware
Evaluation benchmark against GPTZero, Turnitin, and other detectors

License

This model inherits the Apache 2.0 license from Qwen3-8B.

Citation

@misc{aegis-qwen3-8b-humanizer-v1,
  title={Aegis-Qwen3-8B-Humanizer-v1: A DPO-fine-tuned text humanizer},
  author={changcheng967},
  year={2026},
  url={https://huggingface.co/changcheng967/Aegis-Qwen3-8B-Humanizer-v1}
}

Project Aegis — Because students deserve a fair fight against flawed algorithms.

Downloads last month: -

Safetensors

Model size

8B params

Tensor type

BF16

Model tree for changcheng967/Aegis-Qwen3-8B-Humanizer-v1

Base model

Qwen/Qwen3-8B-Base

Finetuned

Qwen/Qwen3-8B

Adapter

(1465)

this model

changcheng967
/

Aegis-Qwen3-8B-Humanizer-v1