t22000t/slay-the-spire-2-card-multimodal-embeddings

Slay the Spire 2: Multimodal Card Embeddings

Joint text+image embeddings for every card in Slay the Spire 2 (Early Access), produced by Qwen/Qwen3-VL-Embedding-2B. One unit-normalized 1024-D vector per card. Mechanically AND visually similar cards land near each other; cards across STS1 and STS2 share the coordinate system.

This is the multimodal-embeddings dataset. For text-only embeddings or the underlying card metadata + portraits, see:

• t22000t/slay-the-spire-2-cards - metadata + features + inline portrait art
• t22000t/slay-the-spire-2-card-embeddings - text-only embeddings via Qwen3-Embedding-0.6B

All three are joinable on id. For the STS1 multimodal counterpart, see t22000t/slay-the-spire-1-card-multimodal-embeddings.

The full bundle (6 datasets across both games + 3 Gradio demos) is in the slaythespire-codex collection.

⚠️ Early Access, content is unstable. Slay the Spire 2 entered Early Access on March 5, 2026. Cards are added, removed, and rebalanced regularly. Always check provenance.json for the snapshot version.

Data Fields

Field	Type	Description
id	string	Stable card identifier - the join key
game	string	Always "sts2"
name	string	Display name
card_text	string	Prettified-JSON document fed to the encoder
has_image	bool	True when the card had a portrait at embed time
multimodal_embedding	list[float32] (1024)	Unit-normalized joint text+image vector

Embedding recipe

Identical to the STS1 multimodal repo so the two are coordinate-compatible:

• Model Qwen/Qwen3-VL-Embedding-2B, frozen.
• Image preprocessing: decode PNG → RGB → resize-with-pad to 512×512 on neutral grey.
• Same task instruction (mechanics primary, art secondary).
• Matryoshka truncation to 1024-D + re-normalize.

See the STS1 multimodal repo for the full instruction wording and rationale.

Loading

from datasets import load_dataset
import numpy as np

ds = load_dataset("t22000t/slay-the-spire-2-card-multimodal-embeddings", split="train")
emb = np.array(ds["multimodal_embedding"], dtype=np.float32)
print(emb.shape)          # (~576, 1024)

Considerations for Using the Data

• Patch drift. STS2 cards turn over in Early Access; vectors only describe the snapshot in provenance.json. Don't compare numerical analyses across snapshots without aligning on game version.
• Lookalike bias. See the STS1 multimodal repo for discussion. Mitigated by the instruction wording but worth evaluating if you observe over-clustering on art alone.
• English only in this snapshot.
• Game IP. Slay the Spire 2 is © Mega Crit. This dataset ships factual reference data + numerical embedding vectors only.

Citation

@dataset{sts2_multimodal_card_embeddings,
  title  = {Slay the Spire 2: Multimodal Card Embeddings},
  author = {timothy22000},
  year   = {2026},
  url    = {https://huggingface.co/datasets/t22000t/slay-the-spire-2-card-multimodal-embeddings},
  note   = {Early Access snapshot; embedded with Qwen3-VL-Embedding-2B; card data via nkhoit/spire-archive; game IP (c) Mega Crit}
}

Licensing

• Dataset: CC BY 4.0
• Pipeline code: MIT
• Embedding model: Apache 2.0
• Game IP: Slay the Spire 2 is © Mega Crit