vltruong01/amazon-all-beauty-led-summarization

Amazon All_Beauty LED Summarization

This dataset contains 2,000 product-level long-context review inputs with synthetic abstractive reference summaries for multi-review summarization.

It is derived from the Amazon Reviews 2023 All_Beauty category and extends the source-only dataset:

vltruong01/amazon-all-beauty-led-reviews

Each example contains multiple selected customer reviews for one Amazon product, one synthetic teacher-generated consensus summary, and quality-control fields used to decide whether the example should be included in supervised training.

Dataset Summary

• Source: Amazon Reviews 2023, All_Beauty category
• Unit of analysis: one product / parent_asin
• Number of rows: 2,000
• Split: train
• Language: English
• Input: product title + multiple selected reviews
• Target: one abstractive product-level consensus summary
• Reference type: synthetic teacher-generated
• Long-context input budget used by the associated LED experiments: 4,096 tokens

The repository intentionally keeps both clean and flagged examples for transparency.

For the associated LED-large training experiment:

2,000 total examples
        ↓
quality / mismatch filtering
        ↓
1,959 clean examples
        ↓
1,567 train
196 validation
196 held-out test

Task

The task is product-level multi-review summarization.

Given several reviews for one product, the model should generate one concise natural-language summary that represents:

• the majority review consensus,
• recurring positive opinions,
• recurring negative opinions,
• meaningful minority complaints,
• product identity,
• important review-supported details.

The numeric product rating is kept as deterministic metadata and is not part of the generated summary target.

Source Construction

The source field comes from a product-level preprocessing pipeline that:

1. cleans individual Amazon reviews,
2. removes duplicate review text,
3. groups reviews by parent_asin,
4. selects informative reviews while attempting to preserve rating diversity and meaningful minority opinions,
5. fits complete review blocks inside the long-context budget,
6. formats the product title and review blocks into one structured source document.

The source-only version of this processing pipeline is available in:

vltruong01/amazon-all-beauty-led-reviews

Input Format

Each source is structured as:

Product: <product title>

[Review 1]
Rating: 5/5
Title: <review title>
Text: <review text>

[Review 2]
Rating: 3/5
Title: <review title>
Text: <review text>

...

[Review N]
Rating: 4/5
Title: <review title>
Text: <review text>

Synthetic Target Construction

target_summary is a synthetic reference summary generated by a GPT-based teacher model.

The teacher is used only to construct reference targets. It is not part of the deployed LED inference pipeline.

The target-generation instructions were designed to encourage summaries that:

• preserve the exact product identity,
• represent the majority consensus,
• retain recurring positive and negative opinions,
• preserve meaningful minority complaints,
• avoid unsupported claims,
• avoid converting isolated opinions into consensus,
• avoid polarity reversal,
• produce one natural paragraph,
• exclude numeric product ratings,
• exclude headings, bullets, and Pros/Cons formatting.

The targets should therefore be interpreted as synthetic supervision, not as independently human-authored ground truth.

Quality Control

The master dataset retains quality-control and mismatch flags rather than deleting flagged rows.

For supervised training, the associated experiment keeps rows satisfying:

target_qc_ok == True
mismatch == False

This filtering yields 1,959 clean examples from the 2,000-row master dataset.

Flag meanings

• target_qc_ok: whether the synthetic target passed target-level quality checks.
• source_title_mismatch: whether a potential inconsistency between the product title and source review content was detected.
• mismatch: final title/source mismatch exclusion flag used by the training pipeline.

Keeping these fields in the public master dataset makes the filtering decision inspectable and reproducible.

Dataset Schema

Column	Type	Description
parent_asin	string	Amazon product-level identifier.
product_title	string	Product title from Amazon metadata.
rating	float	Product-level mean rating computed during preprocessing.
num_reviews	integer	Number of selected reviews represented in the source.
source	string	Long-context structured model input.
target_summary	string	Synthetic teacher-generated abstractive consensus summary.
target_word_count	integer	Number of words in target_summary.
target_qc_ok	boolean	Target quality-control pass/fail flag.
source_title_mismatch	boolean	Detected product-title/source inconsistency flag.
mismatch	boolean	Final mismatch exclusion flag used during model training.

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset(
    "vltruong01/amazon-all-beauty-led-summarization",
    split="train",
)

print(dataset)
print(dataset.column_names)

Recommended Training Filter

def to_bool(value):
    if isinstance(value, bool):
        return value

    if value is None:
        return False

    return str(value).strip().lower() in {
        "true", "1", "yes", "y"
    }


def keep_clean(example):
    return (
        to_bool(example["target_qc_ok"])
        and not to_bool(example["mismatch"])
    )


clean_dataset = dataset.filter(keep_clean)

print("Master rows:", len(dataset))
print("Clean rows:", len(clean_dataset))

Expected clean size for the current 2,000-row version:

Master rows: 2000
Clean rows: 1959

Example Training Fields

For standard supervised summarization:

source_text = example["source"]
target_text = example["target_summary"]

The following metadata fields can be retained for analysis but do not need to be generated by the model:

parent_asin
product_title
rating
num_reviews
target_word_count
target_qc_ok
source_title_mismatch
mismatch

Associated Model

The LED-large model trained with this dataset is available here:

vltruong01/amazon-all-beauty-led-large-summarization

The associated experiment uses:

• allenai/led-large-16384
• maximum source length: 4,096 tokens
• maximum target length: 192 tokens
• full fine-tuning
• 3 epochs
• held-out evaluation after deterministic splitting of the 1,959 clean examples.

For exact training details, use the notebook and experiment artifacts included in the model repository rather than treating the configuration above as part of the dataset itself.

Relationship Between the Three Repositories

Amazon Reviews 2023 — All_Beauty
        ↓
review cleaning / grouping / selection
        ↓
amazon-all-beauty-led-reviews
3,000 source-only product examples
        ↓
synthetic teacher target generation + QC
        ↓
amazon-all-beauty-led-summarization
2,000 source-target examples
        ↓
quality filtering + LED-large fine-tuning
        ↓
amazon-all-beauty-led-large-summarization
fine-tuned model + evaluation artifacts

Intended Use

This dataset is intended for research on:

• multi-review summarization,
• long-context summarization,
• opinion summarization,
• product-level review aggregation,
• synthetic supervision for summarization,
• long-document encoder-decoder fine-tuning,
• factuality and consensus-preservation analysis.

Limitations

• The reference summaries are synthetic teacher-generated targets, not human-written references.
• A high similarity score against target_summary does not guarantee factual correctness.
• Synthetic targets can omit valid minority opinions or overemphasize particular review claims.
• Review selection is constrained by a long-context budget and therefore does not include all available reviews.
• User-generated Amazon reviews may themselves contain subjective, incorrect, exaggerated, or outdated claims.
• The dataset covers only the All_Beauty category and should not be assumed to generalize to all Amazon domains.
• Rows flagged by QC or mismatch fields remain in the master dataset and should normally be filtered before training.

Source Dataset and Citation

This dataset is derived from Amazon Reviews 2023 released by the McAuley Lab.

Original project:

https://amazon-reviews-2023.github.io/

Hugging Face source:

https://huggingface.co/datasets/McAuley-Lab/Amazon-Reviews-2023

If you use this dataset, please cite the original Amazon Reviews 2023 work:

@article{hou2024bridging,
  title={Bridging Language and Items for Retrieval and Recommendation},
  author={Hou, Yupeng and Li, Jiacheng and He, Zhankui and Yan, An and Chen, Xiusi and McAuley, Julian},
  journal={arXiv preprint arXiv:2403.03952},
  year={2024}
}

Acknowledgements

• Source data: McAuley Lab — Amazon Reviews 2023
• Dataset tooling: Hugging Face Datasets
• Synthetic reference construction: GPT-based teacher model
• Associated summarization backbone: Longformer Encoder-Decoder (LED)