Qwen2.5-PromptRefiner-Merged

A fully merged fine-tuned version of Qwen2.5-3B-Instruct trained to transform basic, vague user prompts into high-quality, structured, and effective prompts that get significantly better responses from AI systems. Unlike the adapter-only version, this is a complete standalone model — no base model or PEFT library required.

---

Model Details

Model Description

• Developed by: Kamran (Kamran-56)
• Model type: Causal Language Model (LoRA fine-tuned, fully merged)
• Language(s): English
• License: MIT
• Finetuned from: Qwen/Qwen2.5-3B-Instruct
• Adapter version: Kamran-56/Qwen2.5-3B-PromptRefiner
• Dataset used: Kamran-56/prompt-refinement-dataset

---

Uses

Direct Use

This model takes a basic user-written prompt as input and returns an enhanced, detailed, and well-structured version of the same prompt. It adds role, context, task, format, and constraints to any vague input prompt.

Downstream Use

• Chrome extensions that auto-enhance prompts on AI platforms like ChatGPT, Claude, and Gemini
• API middleware that improves prompts before forwarding to LLMs
• HuggingFace Inference API — can be called directly, no extra setup needed
• Productivity tools that help non-technical users write better prompts

Out-of-Scope Use

• Not designed to answer questions or generate general content
• Not suitable for tasks outside of prompt enhancement
• Not trained on non-English prompts

---

How to Get Started with the Model

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL = "Kamran-56/Qwen2.5-PromptRefiner-Merged"

tokenizer = AutoTokenizer.from_pretrained(MODEL, trust_remote_code=True)

model = AutoModelForCausalLM.from_pretrained(
    MODEL,
    dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)
model.eval()

def enhance_prompt(bad_prompt):
    input_text = f"""<|im_start|>system
You are a world-class prompt engineer with deep expertise across all domains
including coding, writing, business, creativity, and research.

Your job is to transform any basic user prompt into a highly specific,
structured, and effective prompt that will get the best possible response from an AI.

Every enhanced prompt MUST include:
1. A clear ROLE  → "Act as a [specific expert]..."
2. Clear CONTEXT → describe the situation in detail
3. Specific TASK → exactly what needs to be done
4. FORMAT        → how the response should be structured
5. CONSTRAINTS   → tone, length, style, or any boundaries

Rules:
- Keep the original intent and topic
- Be specific, never generic
- Return ONLY the enhanced prompt, nothing else
- No intro, no explanation, no meta-commentary<|im_end|>
<|im_start|>user
{bad_prompt}<|im_end|>
<|im_start|>assistant
"""
    inputs = tokenizer(input_text, return_tensors="pt").to("cuda")
    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=200,
            do_sample=True,
            temperature=0.7,
            top_p=0.9,
            repetition_penalty=1.2,
            pad_token_id=tokenizer.eos_token_id
        )
    full_output = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return full_output.split("assistant")[-1].strip()

print(enhance_prompt("fix my code"))

---

Training Details

Training Data

Fine-tuned on Kamran-56/prompt-refinement-dataset containing 1,561 input→output pairs of basic prompts mapped to high-quality enhanced prompts generated using Llama 3.1 via Groq API.

Training Procedure

Training Hyperparameters

• Training regime: bf16 mixed precision
• Fine-tuning method: LoRA (PEFT) — merged into base model after training
• LoRA rank (r): 8
• LoRA alpha: 16
• Target modules: q_proj, k_proj, v_proj, o_proj
• Learning rate: 5e-5
• Epochs: 3
• Batch size: 4
• Gradient accumulation steps: 4
• Effective batch size: 16
• LR scheduler: Cosine
• Warmup steps: 50
• Max sequence length: 512

Speeds, Sizes, Times

• Hardware: Kaggle P100 GPU (16GB VRAM)
• Training time: ~45 minutes
• Final training loss: 1.24
• Merging method: merge_and_unload() via PEFT

---

Evaluation

Testing Data

Evaluated manually on 12 diverse prompts spanning coding, writing, professional, creative, and general categories.

Results

Average quality rating: 8.5/10 based on manual evaluation.

---

Difference From Adapter Version

---

Citation

@model{qwen25_promptrefiner_merged,
  author    = {Kamran},
  title     = {Qwen2.5-PromptRefiner-Merged},
  year      = {2025},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/Kamran-56/Qwen2.5-PromptRefiner-Merged}
}