qwen2.5-3b-humanizer-merged/README.md

---
language:
- en
license: apache-2.0
base_model: Qwen/Qwen2.5-3B-Instruct
tags:
- text-generation
- humanizer
- rewriter
- qwen2.5
- qlora
- merged
- conversational
pipeline_tag: text-generation
---

# Qwen2.5-3B AI Text Humanizer (Merged)

A fine-tuned version of **Qwen/Qwen2.5-3B-Instruct** that rewrites AI-generated text to sound natural and human-written. This is the **fully merged** model — the LoRA adapter weights have been baked directly into the base model, making it compatible with fast inference engines like **vLLM**.

> ⚠️ This is a **private model**. You need a HuggingFace token with read access to load it.

## Model Details

| Property | Value |
|---|---|
| Base model | `Qwen/Qwen2.5-3B-Instruct` |
| Fine-tuning method | QLoRA (r=64, α=128) |
| Training dataset | `qwertyuiopasdfg/English_humanize` (20k rows) |
| Adapter repo | `arshaan-nazir/qwen2.5-3b-humanizer-qlora` |
| Model type | Causal LM — merged weights |
| Language | English |
| License | Apache 2.0 |

## What It Does

Takes AI-generated text as input and rewrites it to:
- Sound natural and conversational
- Use contractions where appropriate
- Vary sentence length and structure
- Avoid stiff, formal phrasing
- Preserve all original facts

## How to Use

### With Transformers

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

HF_TOKEN = "hf_YOUR_TOKEN_HERE"   # needs read access to this repo

print("Loading tokenizer...")
tokenizer = AutoTokenizer.from_pretrained(
    "arshaan-nazir/qwen2.5-3b-humanizer-merged",
    token=HF_TOKEN,
)

print("Loading model...")
model = AutoModelForCausalLM.from_pretrained(
    "arshaan-nazir/qwen2.5-3b-humanizer-merged",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    token=HF_TOKEN,
)
model.eval()

SYSTEM = """You are a helpful editor.

Rewrite the input so it sounds natural, clear, and conversational while preserving every fact.
Rules:
- Vary sentence length and structure (mix short and long).
- Use contractions where natural.
- Avoid stiff/overly formal phrases.
- No bullet points, headings, or numbered lists.
- Keep paragraphs reasonable (no more than 5 sentences per paragraph).
- Output ONLY the rewritten text (no preamble, no labels).
"""

def humanize(text):
    msgs = [
        {"role": "system", "content": SYSTEM},
        {"role": "user",   "content": f"Rewrite the following so it sounds fully human-written.\n\nTEXT:\n{text}\n\nREWRITE:"}
    ]
    prompt    = tokenizer.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
    input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(model.device)

    with torch.no_grad():
        output = model.generate(
            input_ids,
            max_new_tokens=512,
            do_sample=True,
            temperature=0.75,
            top_p=0.92,
            repetition_penalty=1.08,
            eos_token_id=tokenizer.eos_token_id,
            pad_token_id=tokenizer.eos_token_id,
        )
    return tokenizer.decode(output[0, input_ids.shape[1]:], skip_special_tokens=True)

text = "The utilization of artificial intelligence has resulted in significant advancements."
print(humanize(text))
```

### With vLLM (recommended — fastest)

```python
from vllm import LLM, SamplingParams
from transformers import AutoTokenizer

HF_TOKEN = "hf_YOUR_TOKEN_HERE"

tokenizer = AutoTokenizer.from_pretrained(
    "arshaan-nazir/qwen2.5-3b-humanizer-merged",
    token=HF_TOKEN,
)

llm = LLM(
    model="arshaan-nazir/qwen2.5-3b-humanizer-merged",
    dtype="bfloat16",
    gpu_memory_utilization=0.90,
    max_model_len=4096,
    enforce_eager=True,
)

sampling_params = SamplingParams(
    temperature=0.75,
    top_p=0.92,
    repetition_penalty=1.08,
    max_tokens=1024,
)

SYSTEM = """You are a helpful editor.

Rewrite the input so it sounds natural, clear, and conversational while preserving every fact.
Rules:
- Vary sentence length and structure (mix short and long).
- Use contractions where natural.
- Avoid stiff/overly formal phrases.
- No bullet points, headings, or numbered lists.
- Keep paragraphs reasonable (no more than 5 sentences per paragraph).
- Output ONLY the rewritten text (no preamble, no labels).
"""

text = "Your AI-generated text here..."

msgs = [
    {"role": "system", "content": SYSTEM},
    {"role": "user",   "content": f"Rewrite the following so it sounds fully human-written.\n\nTEXT:\n{text}\n\nREWRITE:"}
]

prompt  = tokenizer.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
outputs = llm.generate([prompt], sampling_params)
print(outputs[0].outputs[0].text.strip())
```

## Recommended Inference Settings

| Parameter | Value |
|---|---|
| `temperature` | 0.75 |
| `top_p` | 0.92 |
| `repetition_penalty` | 1.08 |
| `max_new_tokens` | 512–1024 |

## Training Details

- **Base model:** `Qwen/Qwen2.5-3B-Instruct`
- **Method:** QLoRA with 4-bit NF4 quantization during training
- **LoRA rank:** 64, alpha: 128
- **Target modules:** `q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`
- **Dataset:** `qwertyuiopasdfg/English_humanize` — 20k (input, output) pairs of AI-generated vs human-written text
- **Epochs:** 3
- **Optimizer:** paged_adamw_8bit
- **Learning rate:** 2e-4 with cosine schedule

## Difference from Adapter Repo

| | [`qwen2.5-3b-humanizer-qlora`](https://huggingface.co/arshaan-nazir/qwen2.5-3b-humanizer-qlora) | This repo |
|---|---|---|
| Type | LoRA adapter only | Fully merged model |
| Requires base model | ✅ Yes | ❌ No |
| vLLM compatible | ❌ No | ✅ Yes |
| Size | ~120 MB | ~6 GB |
| Load time | Slower (loads base + adapter) | Faster (single model) |

## Live Demo

Try it at: [arshaan-nazir/ai-text-humanizer](https://huggingface.co/spaces/arshaan-nazir/ai-text-humanizer-1771785713)

## Developer

**arshaan-nazir** — [HuggingFace Profile](https://huggingface.co/arshaan-nazir)