jamesding0302/memgen-annotations

---
task_categories:
- other
configs:
- config_name: AmazonReviews2014-Beauty
data_files:
- split: val
path: AmazonReviews2014-Beauty/val.jsonl
- split: test
path: AmazonReviews2014-Beauty/test.jsonl
- config_name: AmazonReviews2014-Sports_and_Outdoors
data_files:
- split: val
path: AmazonReviews2014-Sports_and_Outdoors/val.jsonl
- split: test
path: AmazonReviews2014-Sports_and_Outdoors/test.jsonl
- config_name: AmazonReviews2023-Industrial_and_Scientific
data_files:
- split: val
path: AmazonReviews2023-Industrial_and_Scientific/val.jsonl
- split: test
path: AmazonReviews2023-Industrial_and_Scientific/test.jsonl
- config_name: AmazonReviews2023-Musical_Instruments
data_files:
- split: val
path: AmazonReviews2023-Musical_Instruments/val.jsonl
- split: test
path: AmazonReviews2023-Musical_Instruments/test.jsonl
- config_name: AmazonReviews2023-Office_Products
data_files:
- split: val
path: AmazonReviews2023-Office_Products/val.jsonl
- split: test
path: AmazonReviews2023-Office_Products/test.jsonl
- config_name: Steam
data_files:
- split: val
path: Steam/val.jsonl
- split: test
path: Steam/test.jsonl
- config_name: Yelp-Yelp_2020
data_files:
- split: val
path: Yelp-Yelp_2020/val.jsonl
- split: test
path: Yelp-Yelp_2020/test.jsonl
---

# MemGen Annotations

This is the annotation dataset for the paper **[How Well Does Generative Recommendation Generalize?](https://huggingface.co/papers/2603.19809)**.

<a href="https://huggingface.co/papers/2603.19809"><img src="https://img.shields.io/badge/Paper-ArXiv-red"></a>
<a href="https://github.com/Jamesding000/MemGen-GR"><img src="https://img.shields.io/badge/Code-GitHub-green"></a>
<a href="https://huggingface.co/jamesding0302/memgen-checkpoints"><img src="https://img.shields.io/badge/Models-Hugging%20Face-blue"></a>

The annotations categorize evaluation instances under the leave-one-out protocol:
- **test** split uses the **last** item in the user history sequence as target,
- **val** split uses the **second-to-last** item as target.

## Columns
- `sample_id`: row index within the split in the original dataset.
- `user_id`: raw user identifier (join key).
- `master`: one of `memorization`, `generalization`, `uncategorized`.
- `subcategories`: list of `{rule, hop}` for fine-grained generalization types.
- `all_labels`: all string labels (e.g., `["generalization", "symmetry_3"]`).

## Load in M&G annotations
```python
from datasets import load_dataset

labels = load_dataset(
"jamesding0302/memgen-annotations",
"AmazonReviews2014-Beauty",
split="test",
)
print(labels[0])
```

## Merge with processed dataset
```python
# 1) Load your processed dataset split (must be aligned with labels by row order)
ds = pipeline.split_datasets["test"]

# 2) Append label columns to the original dataset
ds = (ds
.add_column("master", labels["master"])
.add_column("subcategories", labels["subcategories"])
.add_column("all_labels", labels["all_labels"]))
```