aksern humanizer

a lightweight text humanization model built on llama 3.2 3b.

aksern humanizer is a fine-tuned language model designed to rewrite AI-generated text into more natural, human-like prose while preserving the original meaning.

built on top of meta-llama/Llama-3.2-3B, the model focuses on text transformation rather than general-purpose conversation. its goal is to make generated text sound less formulaic, repetitive, and machine-written without unnecessarily changing the underlying information.

overview

property	value
developer	aksern
model type	causal language model
base model	meta-llama/Llama-3.2-3B
parameter count	3B
primary task	text humanization
architecture	Llama 3.2
intended use	rewriting AI-generated text
inference	local or hosted, depending on deployment

features

• natural rewriting - transforms formulaic AI-generated prose into more natural language.
• meaning preservation - aims to retain the original information and intent.
• lightweight inference - a 3B-parameter model suitable for local deployment on compatible hardware.
• focused behavior - optimized for text transformation instead of unrestricted general-purpose generation.

usage

provide a piece of text and let the model rewrite it while preserving its meaning. example code:

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

path = "aksern/humanizer"
text = "In today's rapidly evolving technological landscape, it is important to note that artificial intelligence plays a significant role in various aspects of our daily lives."
model = AutoModelForCausalLM.from_pretrained(path, dtype=torch.bfloat16, device_map="auto")
tok = AutoTokenizer.from_pretrained(path)
inputs = tok.apply_chat_template(
    [{"role": "user", "content": text}],
    add_generation_prompt=False, tokenize=True, return_dict=True, return_tensors="pt",
).to(model.device)
with torch.no_grad():
    outputs = model.generate(
        **inputs, max_new_tokens=120, do_sample=True,
        temperature=0.6, top_p=0.9, pad_token_id=tok.eos_token_id,
    )
out_ids = outputs[0][inputs["input_ids"].shape[-1]:]
if tok.eos_token_id not in out_ids.tolist():
    print("WARNING: generation hit max_new_tokens, not EOS")
print("--- GENERATED ---")
print(repr(tok.decode(out_ids, skip_special_tokens=True)))

example

input

In today's rapidly evolving technological landscape, it is important to note that artificial intelligence plays a significant role in various aspects of our daily lives.

humanized output

In this fast paced technological world, we need to realize that artificial intelligence is a major factor in many aspects of our lives.

intended use

• rewriting AI-generated drafts into more natural prose
• reducing repetitive or overly formulaic phrasing
• improving the readability and flow of generated text
• experimenting with lightweight text-humanization models

limitations

• Humanization quality depends on the input text, language, and generation settings.
• The model may not perform equally well across languages, domains, and writing styles.
• More natural-sounding text does not necessarily mean more accurate text.
• Humanization and AI-text detection are different tasks. The model is not guaranteed to bypass AI detectors.

acknowledgements

Built by aksern.

models that run where you are.