---
license: apache-2.0
base_model: unsloth/Qwen3-4B-Instruct-2507
tags:
- qwen
- unsloth
- quantized
- q4_k_m
- gguf
- fine-tuned
- text-rewriting
- text-generation
- human-like-writing
language:
- en
datasets:
- custom-jsonl-rewriting
library_name: gguf
pipeline_tag: text-generation
---

# Qwen3-4B-Instruct Text Rewriter (GGUF Format)

### Example Notebook
🔗Kaggle notebook [here](https://www.kaggle.com/code/soumyojitsen2004/example-code)

## Model Description
This model is a fine-tuned version of **Qwen3-4B-Instruct-2507**, optimized specifically to rewrite AI-generated text into a more natural, human-like style. It was trained to reduce the "robotic" tone often found in LLM outputs while maintaining the original meaning and factual accuracy.

- **Base Model:** [unsloth/Qwen3-4B-Instruct-2507](https://huggingface.co/unsloth/Qwen3-4B-Instruct-2507) (merged from `unsloth/Qwen3-4B-Instruct-2507-bnb-4bit`)
- **Fine-tuned by:** [Heatw4ve](https://huggingface.co/Heatw4ve)
- **Fine-tuning Tool:** [Unsloth](https://github.com/unslothai/unsloth)
- **Quantization format:** GGUF (Quantized to Q4_K_M)
- **Training framework:** Unsloth (LoRA-based SFT)  
- **Language:** English
- **Intended usage:** Local inference, research, and experimentation  

## Training Details
The model was fine-tuned using LoRA (Low-Rank Adaptation) on a custom dataset of 28,000+ pairs of AI vs. Human-rewritten text. 

### Training Configuration 
The model was fine-tuned with the following parameters:
- **Epochs:** 2
- **Learning Rate:** 5e-5
- **Optimizer:** AdamW 8-bit
- **LR Scheduler:** Cosine
- **Warmup Steps:** 100
- **Context Length:** 512 tokens
- **LoRA Config:** - Rank (r): 64
  - Alpha: 128
  - Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
- **Effective Batch Size:** 128 (32 per device * 4 grad accumulation)

## Usage
This is a GGUF version, compatible with **LM Studio**, **Ollama**, and **llama.cpp**. Works best with inputs of length 15-150 words.

### System Prompt
To achieve the intended rewriting style, use the following system prompt:
> You are a helpful assistant that rewrites AI-toned text into natural, human-like writing.

### User Prompt Format
```text
Rewrite the following text to sound like a real human wrote it:

[INSERT TEXT HERE]
```

### Example Code
```code
from llama_cpp import Llama
llm = Llama.from_pretrained(
  repo_id="Heatw4ve/Qwen3-4B-Instruct-2507-rewriter-gguf",
  filename="Qwen3-4B.gguf",
  n_ctx=4096,
  n_gpu_layers=16,
  n_threads=4,
  verbose=False,
  chat_format="chatml",
)

text='''
William Shakespeare, the master of human insight, gifted us with a timeless observation that cuts to the heart of intellect and humility: "The fool doth think he is wise, but the wise man knows himself to be a fool." Uttered by the character Touchstone in As You Like It, this seemingly paradoxical statement is far more than a witty quip; it's a profound commentary on self-awareness, the nature of true wisdom, and the perpetual quest for knowledge.
'''

messages = [
  {"role": "system", "content": "You are a helpful assistant that rewrites AI-toned text into natural, human-like writing."},
  {"role": "user", "content": f"Rewrite the following text to sound like a real human wrote it:\n\n{text}"}
]

output = llm.create_chat_completion(
  messages=messages,
  max_tokens=512,
  temperature=random.uniform(0.9, 1.2),
  top_p=random.uniform(0.87, 0.96),
  seed=random.randint(2, 2**32),
  stop=["<|im_end|>", "<|endoftext|>"]
)

print(output["choices"][0]["message"]["content"].strip())
```

🔗[Connect](https://www.linkedin.com/in/soumyojit-sen/) with me on LinkedIn