language:
- en
license: mit
configs:
- config_name: cards
data_files:
- split: train
path: mtg_embeddings.parquet
default: true
Dataset Card for Dataset Name
Text embeddings of all Magic: The Gathering card until Aetherdrift (2024-02-14). The text embeddings are centered around card mechanics (i.e. no flavor text/card art embeddings) in order to identify similar cards mathematically.
This dataset also includes zero-mean-centered 2D UMAP coordinates for all the cards, in columns x_2d and y_2d.
Dataset Details
How The Embeddings Were Created
Using the data exports from MTGJSON, the data was deduped by card name to the most recent card printing and aggregated as a prettified JSON object string (including indentation, which is relevant to embedding quality).
An example card input string:
{
"name": "Marhault Elsdragon",
"manaCost": "{3}{R}{R}{G}",
"type": "Legendary Creature — Elf Warrior",
"text": "Rampage 1 (Whenever this creature becomes blocked, it gets +1/+1 until end of turn for each creature blocking it beyond the first.)",
"power": "4",
"toughness": "6",
"rarities": [
"uncommon"
],
"sets": [
"LEG",
"BCHR",
"CHR",
"ME3"
]
}
This JSON string was then encoded and unit-normalized using the Alibaba-NLP/gte-modernbert-base text embedding model.
JSON schema definitions:
name: Card namemanaCost: Card mana costtype: Card type (full textline)text: Gatherer card text. Exact matches to card names within card text are replaced with~for consistency.power: Card power. Omitted if N/A.toughness: Card toughness. Omitted if N/A.loyalty: Card loyalty. Omitted if N/A.rarities: List of all card rarities for all printings of the card. Rarities are listed in this order:["common", "uncommon", "rare", "mythic", "bonus", "special"]sets: List of all sets the card was printed in. Sets are listed in chronological order of printing. Basic lands have a special case of[*].
It took 1:17 to encode all ~33k cards using a preemptible L4 GPU on Google Cloud Platform (g2-standard-4) at ~$0.28/hour, costing < $0.01 overall.