anonymousemnlpauthor/games-asde-qwen25-3b-lora

Games ASDE Preference Extractor

This repository contains a LoRA adapter for an aspect-level preference extractor trained for the Games domain.

The model is designed to extract structured preference signals from review sentences. Given a review sentence, it identifies which product aspect is discussed, what descriptor expresses the user's opinion, the sentiment polarity, and the supporting evidence span from the original sentence.

This repository contains the adapter weights only. The base model must be loaded separately.

Model Details

• Model type: PEFT LoRA adapter for causal language modeling
• Base model: Qwen/Qwen2.5-3B-Instruct
• Task: Aspect-level preference extraction from review text
• Domain: Games / video game reviews
• Language: English
• Output format: Structured JSON-style extraction
• Adapter framework: PEFT
• Fine-tuning method: Supervised fine-tuning with LoRA
• Shared by: anonymousemnlpauthor

Intended Task

The model extracts aspect-level preference signals from user review sentences.

Each extracted mention follows the structure below:

{
  "a": "Aspect category",
  "d": "Descriptor phrase",
  "s": 1,
  "e": {
    "text": "Evidence span copied from the sentence",
    "char_start": 0,
    "char_end": 10
  }
}

Field Description

Field	Meaning
a	Aspect category discussed in the sentence
d	Descriptor phrase summarizing the user's opinion or experience
s	Sentiment polarity: 1 positive, 0 neutral/mixed, -1 negative
e.text	Evidence span copied from the input sentence
e.char_start	Start character offset of the evidence span
e.char_end	End character offset of the evidence span

Example Use Case

Input sentence:

The controls are responsive, but the story feels too short.

Expected style of output:

{
  "labels": [
    {
      "a": "Control responsiveness",
      "d": "responsive",
      "s": 1,
      "e": {
        "text": "controls are responsive",
        "char_start": 4,
        "char_end": 27
      }
    },
    {
      "a": "Narrative and immersion",
      "d": "story feels too short",
      "s": -1,
      "e": {
        "text": "story feels too short",
        "char_start": 37,
        "char_end": 58
      }
    }
  ]
}

The exact output depends on the prompt format used during inference.

How to Load

Install the required packages:

pip install -U transformers peft accelerate safetensors

Load the adapter with the base model:

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

base_model = "Qwen/Qwen2.5-3B-Instruct"
adapter_model = "anonymousemnlpauthor/games-asde-qwen25-3b-lora"

tokenizer = AutoTokenizer.from_pretrained(adapter_model)

model = AutoModelForCausalLM.from_pretrained(
    base_model,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

model = PeftModel.from_pretrained(model, adapter_model)
model.eval()

print("Adapter loaded successfully.")

Minimal Inference Example

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

base_model = "Qwen/Qwen2.5-3B-Instruct"
adapter_model = "anonymousemnlpauthor/games-asde-qwen25-3b-lora"

tokenizer = AutoTokenizer.from_pretrained(adapter_model)
model = AutoModelForCausalLM.from_pretrained(
    base_model,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
model = PeftModel.from_pretrained(model, adapter_model)
model.eval()

sentence = "The controls are responsive, but the story feels too short."

prompt = f"""You are an aspect-level preference extractor.
Extract aspect-level preference mentions from the review sentence.
Return a JSON object with a key named "labels".

Sentence:
{sentence}
"""

messages = [
    {"role": "user", "content": prompt}
]

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

with torch.no_grad():
    outputs = model.generate(
        inputs,
        max_new_tokens=256,
        do_sample=False
    )

decoded = tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True)
print(decoded)

Intended Use

This model is intended for research on personalized recommendation explanation generation, user preference modeling, and aspect-level review understanding.

Possible uses include:

• Extracting aspect-level user preferences from historical reviews
• Building multi-aspect user profiles
• Constructing aspect-level evidence for personalized review generation
• Analyzing which product aspects users discuss positively or negatively

Out-of-Scope Use

This model is not intended for:

• Making high-stakes decisions about users
• Inferring sensitive personal attributes
• Producing factual claims without downstream verification
• General-purpose chat or instruction following
• Domains substantially different from game/video game reviews without additional validation

Limitations

• The model is trained for the Games review domain and may not generalize well to other domains.
• The model may produce invalid JSON, incomplete spans, or incorrect character offsets.
• The model may assign an aspect category even when the sentence is ambiguous.
• The extracted sentiment reflects the local review sentence, not necessarily the user's global preference.
• Evidence spans should be checked before being used in downstream systems that require strict grounding.

Training Data

The adapter was fine-tuned on a Games-domain review corpus annotated for aspect-level preference extraction. Each training example contains review sentences and structured labels consisting of aspect, descriptor, sentiment polarity, and evidence span.

The training data is not included in this repository.

Training Procedure

The model was trained by supervised fine-tuning using LoRA on top of Qwen/Qwen2.5-3B-Instruct.

Known framework version from the saved adapter:

• PEFT: 0.18.1

Evaluation

Evaluation results are not included in this model card. Users should evaluate the model on their own target domain and prompt format before deployment.

Recommended checks include:

• JSON validity
• Aspect classification accuracy
• Descriptor quality
• Sentiment polarity accuracy
• Evidence span exact match or overlap
• Character offset correctness

Citation

If you use this model, please cite the associated paper or repository when it becomes available.

@misc{anonymous2026gamesasde,
  title        = {Games ASDE Preference Extractor},
  author       = {Anonymous},
  year         = {2026},
  howpublished = {Hugging Face model repository}
}

Contact

For questions, please contact the repository owner.