Qwen2.5-7B-Ecom-Refiner is a specialized LoRA adapter designed to enhance user-generated product descriptions for E-commerce platforms. It transforms raw, unstructured facts into professional, engaging, and well-structured sales copy.

📌 Overview

Project Goal: Improve conversion rates and user experience (UX) on trading platforms by automating the creation of high-quality product listings.
Base Model: Qwen/Qwen2.5-7B-Instruct
Language: Russian (RU)
Training Method: QLoRA (4-bit quantization)

🛠 Training Methodology

Dataset & Knowledge Distillation

The model was trained on a curated dataset of 1,500 examples created through knowledge distillation:

Teacher Model: DeepSeek V3.
Process: The teacher model generated "gold standard" descriptions based on raw, unstructured data from open E-commerce archives.
Focus: Maintaining factual accuracy while improving readability and structure.

Technical Stack

Libraries: TRL, Transformers, PEFT.
Hardware: Optimized for training on 2x Tesla T4 GPUs.
Quantization: 4-bit (QLoRA) to ensure efficient performance within 16-32 GB VRAM.

📊 Evaluation & Results

The model was evaluated using both semantic similarity and structural adherence metrics:

Metric	Value	Description
BERTScore (F1)	0.79	High factual consistency and minimal hallucinations during paraphrasing.
Structure Adherence	91%	Success rate in following the required format (bullet points, clear accents).

🚀 Quick Start (Inference)

To use this adapter, you need to load the base Qwen2.5-7B-Instruct model and apply the LoRA weights. Below is an example based on the inference.py script:

python

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

model_id = "Qwen/Qwen2.5-7B-Instruct"
adapter_id = "yuriy-magus/Qwen2.5-7B-Ecom-Refiner"
device = "cuda" if torch.cuda.is_available() else "cpu"

# 1. Load Tokenizer and Base Model
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16 if device == "cuda" else torch.float32,
    device_map="auto" if device == "cuda" else None,
    trust_remote_code=True
)

# 2. Load Adapter
model = PeftModel.from_pretrained(model, adapter_id)
model.to(device).eval()

# 3. Prepare Prompt (as used in training)
category = "Личные вещи / Одежда, обувь, аксессуары"
title = "Кеды pataugas"
original_desc = "Кеды фиpмы PATAUGAS. Кеды пoлнocтью кoжaные, пoдoшвa пpoшитa, мoлнии paбoчие. Сocтoяние хopoшее."
prompt = f"""### Instruction:
Отредактируй описание товара для Авито. Будь грамотным, добавь в текст структуру и привлекательность, а также строго придерживайся фактов из исходного текста.

### Context:
Категория: {category}
Товар: {title}

### Original Description:
{original_desc}

### Improved Description:
"""

# 4. Generate
inputs = tokenizer(prompt, return_tensors="pt").to(device)
with torch.no_grad():
    outputs = model.generate(
        **inputs, 
        max_new_tokens=200, 
        temperature=0.7, 
        top_p=0.9,
        do_sample=True,
        pad_token_id=tokenizer.eos_token_id
    )

result = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(result.split("### Improved Description:")[-1].strip())