This model is a fine-tuned version of Qwen/Qwen3.5-2B using QLoRA (Low-Rank Adaptation) for Aspect-Based Sentiment Analysis (ABSA) on multilingualnd bilingual restaurant reviews.

Developed as part of the Large Language Models course at the Universitat Politècnica de Catalunya (UPC) - School of Informatics of Barcelona (FIB), under the supervision of Dr. Lluís Màrquez Villodre.

Model Description

The model is specialized in constrained generation for structured information extraction. Given a restaurant review (in English or Spanish), it enforces a strict JSON output matching a predefined schema of 11 valid aspect categories and 4 valid polarities, effectively bypassing the need for separate heuristic parsers.

Allowed Schema

Aspect Categories: restaurant_general, food_quality, service, ambience, food_style_options, food_prices, restaurant_prices, drinks_quality, drinks_style_options, location, drinks_prices.
Sentiment Polarities: positive, negative, neutral, conflict.
Developed by: Roger Baiges Trilla & Guillem Olivart Garrofé
Language(s): Bilingual
License: Apache-2.0
Finetuned from model: Qwen/Qwen3.5-2B

Experimental Results & Performance

The model was iteratively evaluated on the devel.json dataset across the four mandatory stages of the project. The QLoRA fine-tuning stage proved to be the most robust approach to combat dataset imbalance, yielding the highest scores across all macro and micro metrics.

Main Development-Set Results

Configuration	Precision (Macro)	F1 (Macro)	F1 (Micro)
Empty baseline	0.0%	0.0%	0.0%
Top 3 majority baseline	60.4%	55.6%	56.2%
Best Zero-Shot Sweep	74.6%	67.9%	68.3%
Best Few-Shot	79.6%	73.9%	73.9%
Best Full LoRA Checkpoint	86.4%	84.4%	86.4%
Best QLoRA Checkpoint (This Model)	87.4%	85.3%	87.3%

Key Takeaway: Fine-tuning permanently alters the decision boundaries of the small language model (SLM), providing a massive +17.4% F1-macro boost over the optimized zero-shot baseline and proving that 4-bit quantization did not degrade extraction capabilities.

Training Prompt Template

To guarantee perfect reproducibility and maintain the 85.3% F1-macro performance, this model requires a highly specific constraint-enforcement prompt.

The complete schema, guidelines, and instructions are explicitly stored in the repository within the absa_prompt.json file, separated into system and user roles to maximize Qwen's instruction-following capabilities.

How to Get Started (Inference Example)

This model is a PEFT adapter. The example below automatically fetches the official absa_prompt.json prompt from this Hugging Face repository at runtime, formats the text, and runs the inference:

python

from peft import AutoPeftModelForCausalLM
from transformers import AutoTokenizer
from huggingface_hub import hf_hub_download
import json
import torch

repo_id = "guillemolivart/qwen-2b-absa-qlora"

# 1. Download and load the prompt template from the repository files
prompt_file_path = hf_hub_download(repo_id=repo_id, filename="absa_prompt.json")
with open(prompt_file_path, "r", encoding="utf-8") as f:
    prompt_data = json.load(f)

# 2. Load the model and tokenizer in quantized 4-bit for high efficiency
model = AutoPeftModelForCausalLM.from_pretrained(
    repo_id,
    load_in_4bit=True,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(repo_id)

# 3. Define your custom multilingual review
review_text = "La comida buenísima, especialmente la carne, de precio correcto, pero los camareros tardaron una eternidad en llevarnos la cuenta."
review_language = "es"  # Supports "es" and "en"

# Format the user part of the prompt
user_content = prompt_data["user"].format(language=review_language, text=review_text)

# Build the ChatML message array using the separated system and user roles
messages = [
    {"role": "system", "content": prompt_data["system"]},
    {"role": "user", "content": user_content}
]

inputs = tokenizer.apply_chat_template(
    messages, 
    tokenize=True, 
    add_generation_prompt=True, 
    return_tensors="pt"
).to("cuda")

outputs = model.generate(
    inputs, 
    max_new_tokens=256, 
    temperature=0.1, 
    do_sample=False
)

# 4. Decode response skipping the prompt system structure
generated_tokens = outputs[0][len(inputs[0]):]
response = tokenizer.decode(generated_tokens, skip_special_tokens=True)

print(response)
# Expected Output: {"food_quality": "positive", "food_prices": "neutral", "service": "negative"}