aiml8726737/personal-prompt-engineer-qwen1.5b

Personal Prompt Engineer (Qwen2.5-1.5B LoRA)

Personal Prompt Engineer is a QLoRA fine-tuned LoRA adapter built on Qwen2.5-1.5B-Instruct. It specializes in transforming short, vague, or incomplete prompts into structured, detailed, and execution-ready prompts while preserving the user's original intent.

Unlike a general chatbot, this model is designed to rewrite prompts, not answer them.

---

Model Details

Property	Value
Base Model	Qwen2.5-1.5B-Instruct
Fine-tuning Method	QLoRA
Adapter Type	LoRA
Quantization	4-bit NF4
Framework	Transformers + PEFT
Primary Task	Prompt Rewriting

---

Intended Use

The model is designed for:

• Prompt Engineering
• AI Assistants
• RAG Pipelines
• Workflow Automation
• Coding Assistants
• Content Generation
• Agentic AI Systems

Typical workflow:

User Prompt
      ↓
Personal Prompt Engineer
      ↓
Optimized Prompt
      ↓
GPT / Claude / Gemini / Llama / Qwen

---

Example

Input

make a react app with dark mode

Output

Create a modern React application with a responsive user interface that includes a dark/light mode toggle.

Requirements:
- Use React functional components.
- Implement state management using Context API.
- Persist theme preference in localStorage.
- Include accessible UI components.
- Follow clean project architecture.
- Provide complete source code with explanations.

---

Training Dataset

The adapter was fine-tuned on approximately 3,030 curated prompt rewriting examples spanning:

• Software Development
• Business
• Marketing
• Creative Writing
• Education
• Science
• Productivity
• General AI Prompting

Each example consists of:

• Instruction
• Draft Prompt
• Rewritten Prompt

---

Training Configuration

Hyperparameter	Value
LoRA Rank	16
LoRA Alpha	32
LoRA Dropout	0
Optimizer	paged_adamw_8bit
Learning Rate	2e-4
Scheduler	Cosine
Epochs	2
Sequence Length	512
Batch Size	4
Gradient Accumulation	4
Effective Batch Size	16

---

Evaluation

Epoch	Training Loss	Validation Loss	Mean Token Accuracy
1	1.1990	1.1171	74.08%
2	0.9854	1.0420	76.45%

The validation loss consistently decreased during training, indicating good convergence without obvious signs of overfitting.

---

Installation

pip install transformers peft accelerate bitsandbytes

---

Usage

import warnings warnings.filterwarnings("ignore")

Bypass torchao metadata version check across all environments

import importlib.metadata _orig_version = importlib.metadata.version def _mock_version(package_name, *args, **kwargs): if package_name == "torchao": return "1.0.0" return _orig_version(package_name, *args, **kwargs) importlib.metadata.version = _mock_version

import torch from transformers import pipeline, AutoModelForCausalLM, AutoTokenizer from peft import PeftModel

1. Hugging Face Repositories

BASE_MODEL = "Qwen/Qwen2.5-1.5B-Instruct" LORA_ADAPTER = "aiml8726737/personal-prompt-engineer-qwen1.5b"

2. Device Selection

device = "cuda" if torch.cuda.is_available() else ("mps" if hasattr(torch.backends, "mps") and torch.backends.mps.is_available() else "cpu") dtype = torch.float16 if device in ["cuda", "mps"] else torch.float32

print(f"Loading Hugging Face Pipeline on {device.upper()}...")

3. Load Tokenizer & Base Model

tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL) tokenizer.pad_token_id = tokenizer.eos_token_id

base_model = AutoModelForCausalLM.from_pretrained( BASE_MODEL, torch_dtype=dtype, device_map="auto" if device == "cuda" else None ) if device != "cuda": base_model = base_model.to(device)

4. Load Fine-Tuned Adapter

model = PeftModel.from_pretrained(base_model, LORA_ADAPTER)

5. Create Standard Hugging Face Text-Generation Pipeline

generator = pipeline( "text-generation", model=model, tokenizer=tokenizer, device_map="auto" if device == "cuda" else None )

SYSTEM_PROMPT = "You are an expert Personal Prompt Engineer. Your task is to rewrite vague user prompts into professional, execution-ready prompts."

6. Prompt Rewriter Function using Hugging Face Pipeline

def rewrite_prompt(draft_prompt: str) -> str: messages = [ {"role": "system", "content": SYSTEM_PROMPT}, {"role": "user", "content": f"Rewrite the following draft prompt into a professional, execution-ready prompt.\n\nDraft Prompt:\n<<>> {draft_prompt} <<>>"} ]

# Hugging Face Pipeline Execution
result = generator(
    messages,
    max_new_tokens=300,
    temperature=0.7,
    do_sample=True,
    return_full_text=False
)

return result[0]["generated_text"]

7. Test Example

if name == "main": draft = "create a landing page for an AI agent app" print("\n" + "="*70) print("DRAFT PROMPT:") print(draft) print("="*70) print("\nREWRITTEN BY HUGGINGFACE PIPELINE:") print(rewrite_prompt(draft)) print("="*70)

Limitations

• Optimized for prompt rewriting rather than general question answering.
• Performance depends on the quality and diversity of the training data.
• May not generalize well to highly specialized domains absent from the training set.

---

License

This LoRA adapter is released under the Apache 2.0 License, consistent with the license of the base Qwen2.5 model.

---

Citation

@misc{personalpromptengineer2026,
  title={Personal Prompt Engineer: QLoRA Fine-tuning for Prompt Rewriting},
  author={Yashvardhan Agrawal},
  year={2026},
  howpublished={Hugging Face Model Hub}
}