t22000t/optcg-en-card-embeddings

OPTCG English Card Embeddings

Text embeddings for the One Piece Card Game (English) generated with Qwen/Qwen3-Embedding-0.6B, plus UMAP 2-D coordinates for visualization. Joins to the sibling t22000t/optcg-en-cards dataset on id.

Generated by the open-source optcg-cards pipeline. Refreshed monthly.

Dataset summary

• One row per card. Same rows as the cards repo, plus three columns:
  • embedding: list of floats. L2-normalized; cosine similarity = dot product. Default dim 1024 (full Qwen3-Embedding-0.6B output).
  • umap_x, umap_y: float UMAP 2-D coordinates.
• Vectors are produced by encoding a structured "card document" — a prettified JSON of mechanics-relevant fields, with the card's own name substituted out and a task instruction prepended. See DOCUMENTATION.md §6 for the full recipe.

Quick start

import pandas as pd, numpy as np

df = pd.read_parquet("hf://datasets/t22000t/optcg-en-card-embeddings/cards_with_embeddings.parquet")

# Pack into a (n, d) matrix
matrix = np.stack(df["embedding"].apply(np.asarray).values).astype("float32")

# Top-k nearest neighbors of card index 0 (cosine; vectors pre-normalized)
query = matrix[0]
scores = matrix @ query
top = scores.argsort()[::-1][:10]
print(df.iloc[top][["id", "name", "card_type", "colors"]])

Joining with the cards repo

The two datasets share id as a primary key:

import pandas as pd
cards = pd.read_parquet("hf://datasets/t22000t/optcg-en-cards/cards.parquet")
emb   = pd.read_parquet("hf://datasets/t22000t/optcg-en-card-embeddings/cards_with_embeddings.parquet")
joined = cards.merge(emb[["id", "embedding"]], on="id", how="left")

(In practice the embeddings repo carries the full card columns too, so a join is only useful if you want columns this snapshot doesn't duplicate.)

Embedding details

• Model: Qwen/Qwen3-Embedding-0.6B.
• Dim: 1024 (default; configurable via Matryoshka truncation in the source pipeline).
• Normalization: L2 (cosine ≡ dot product).
• Document strategy: prettified JSON of name, card_type, colors, cost, power, counter, life, attribute, family, effect_text, trigger_text, rarity. The card's own name is substituted with ~ in effect_text / trigger_text.
• Task instruction: a fixed string prepended to each document telling the model to optimize for OPTCG mechanic similarity. The exact string is captured in provenance.json:embed.task_instruction for reproducibility.

Reproducing

The full pipeline is open-source. To recompute these embeddings:

pip install 'optcg-cards[embed]'
optcg-cards fetch en
optcg-cards embed en
optcg-cards visualize en

Or to encode a custom query against this dataset:

from sentence_transformers import SentenceTransformer
import json

prov = json.loads(open("provenance.json").read())
model = SentenceTransformer(prov["embed"]["model_id"])
text = prov["embed"]["task_instruction"].format(card_document="your query here")
query_vec = model.encode([text], normalize_embeddings=True)[0]

Provenance

provenance.json ships in this repo. It carries both the fetch block (which upstream snapshot was used) and the embed block (model, embedding dim, task instruction, embed timestamp). The instruction is stored verbatim — you need it to encode queries that are comparable to the stored vectors.

Intellectual property

This embeddings dataset is a numeric transformation of card-text strings. It does not contain card art. See the sibling cards dataset README for the full IP notice and the takedown contact.

Citation

@misc{optcg_en_card_embeddings,
  title  = {OPTCG English Card Embeddings},
  author = {optcg-cards contributors},
  year   = {2026},
  url    = {https://huggingface.co/datasets/t22000t/optcg-en-card-embeddings},
  note   = {Qwen3-Embedding; sister of t22000t/optcg-en-cards}
}

License

CC BY 4.0. See LICENSE-DATA in the source repository for the full license text and IP notice.