t22000t/slay-the-spire-2-cards

Slay the Spire 2: Cards

A normalized dataset of every card in Slay the Spire 2 (Early Access), with derived feature columns including STS2-specific Orb, Forge, and Soul mechanics. Collected for ML/DL training: load with datasets.load_dataset(...) and feed straight into a card-text classifier, deckbuilder simulator, or design-analysis model.

This is the cards dataset. Companion datasets, all joinable on id:

• t22000t/slay-the-spire-2-card-embeddings - 1024-D text embeddings via Qwen/Qwen3-Embedding-0.6B
• t22000t/slay-the-spire-2-card-multimodal-embeddings - joint text+image embeddings via Qwen/Qwen3-VL-Embedding-2B

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

For Slay the Spire 1, see t22000t/slay-the-spire-1-cards.

⚠️ 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 the sts_game_version field in provenance.json before drawing comparisons across snapshots.

Why split from embeddings?

Card-text classification, design analysis, modding, and deckbuilder simulators don't need the embedding column. Splitting also lets the embedding repo be re-versioned independently when the embedding model changes, without touching the cards repo's commit history. Both repos are joinable on id.

Data Fields

Columns appear below in the order they're stored on disk. id and the portrait columns lead the row so the HF dataset viewer surfaces a thumbnail before metadata text.

Identifier + portrait

Field	Type	Description
id	string	Stable card identifier, the join key to the embeddings, multimodal-embeddings, and derived datasets
image	HF Image feature (struct<bytes, path>)	Card portrait art. Decoded to a PIL Image automatically by datasets.load_dataset(); via pd.read_parquet you'll get a dict ({"bytes": ..., "path": null}). Every STS2 card in this snapshot ships a portrait.
image_resolution	string	Resolution variant of the source asset.

Card metadata

Field	Type	Description
game	string	Always "sts2"
name	string	Display name
type	string	One of Attack, Skill, Power, Status, Curse, Quest
rarity	string	One of Basic, Common, Uncommon, Rare, Special, Curse, Status, Ancient, Event, Quest, Token (Title Case)
color	string	Character class or pseudo-class (lowercase, as emitted by the upstream API): ironclad, silent, defect, necrobinder, regent, colorless, curse, event, quest, status, token. New character classes may appear with future patches
cost	string	Energy cost ("0", "1", "X", "-" for unplayable)
description	string	Card text, base form
description_upgraded	string	Card text after upgrade (+)
keywords	string (JSON list)	Declared keywords
raw_json	string (JSON object)	Full original payload (preserves Orbs/Forge/Souls/Enchantments structure)

Generic feature columns

Numeric columns surface as double on the wire (pandas casts ints with nulls to float); semantically they're integer counts.

Field	Type	Description
damage	float | null	Primary damage value (e.g. "Deal 6 damage" → 6)
damage_upgraded	float | null	Same for the upgraded form
block	float | null	Primary block value
block_upgraded	float | null	Same for the upgraded form
targets_all_enemies	bool	True if the card hits ALL enemies (AOE)
status_effects_applied	string (JSON list)	[{effect, count}] for Vulnerable/Weak/Frail/etc.
mechanics	string (JSON list)	Mechanic keywords (declared + inferred from text)

STS2-specific feature columns

Field	Type	Description
orbs_channeled	string (JSON list)	[{type, count}] for Orbs the card channels
orbs_referenced	string (JSON list)	Orb types mentioned in the description, channeled or not
forge_value	float | null	Forge value if the card has a "Forge N" clause
souls_added	float | null	Number of Souls added (Necrobinder mechanic)

Card art

Each row carries the in-game card portrait as a PIL-decodable image. The HuggingFace dataset viewer renders thumbnails inline.

• Source. Card portraits are extracted from a local Steam install of Slay the Spire 2 (sts2.pck, expanded with GDRE Tools). Game files are not redistributed by this pipeline; extraction happens on the maintainer's machine.
• Drift. STS2 art changes alongside cards in Early Access. The provenance.json records art.source_file_sha256 and art.extracted_at so consumers can detect rebuilt portraits.
• Why this is large. The cards Parquet grows from ~250 KB to ~80 MB once portraits are inlined. If you don't need the bytes, use streaming=True.
• Attribution. Slay the Spire 2 art is © Mega Crit. Mega Crit has publicly blessed redistribution of fan-extracted card art for tooling and mods. This dataset includes a takedown clause regardless: if Mega Crit objects, the image column will be removed.

Loading

from datasets import load_dataset

cards = load_dataset("t22000t/slay-the-spire-2-cards", split="train")
print(cards[0]["name"], "→", cards[0]["image"])
# <name> → <PIL.PngImagePlugin.PngImageFile image mode=RGBA size=...>

Filtering on STS2 mechanics

import json
import pandas as pd

df = cards.to_pandas()

# All cards that channel a Lightning orb
def channels_lightning(row):
    return any(o["type"] == "Lightning" for o in json.loads(row))

mask = df["orbs_channeled"].apply(channels_lightning)
print(df.loc[mask, ["name", "description"]])

Joining with the embeddings dataset

from datasets import load_dataset

cards = load_dataset("t22000t/slay-the-spire-2-cards", split="train").to_pandas()
embs  = load_dataset("t22000t/slay-the-spire-2-card-embeddings", split="train").to_pandas()

df = cards.merge(embs[["id", "embedding"]], on="id", how="inner")

Considerations for Using the Data

Early Access Drift, the most important caveat

STS2 cards change frequently. A snapshot from one week may contain cards that no longer exist, missing cards added since, or rebalanced versions of existing cards. Always read provenance.json to know which patch this snapshot reflects. Don't compare numerical analyses across snapshots without aligning on game version.

Discussion of Biases

The derived feature columns are extracted via best-effort regex tuned for English STS2 phrasing. STS2-specific mechanics (Orbs, Forge, Souls) are extracted via name lookups against an explicit allowlist; new mechanics introduced in future patches won't appear in derived columns until the regex is updated. The full original text is always available in description and raw_json for re-extraction.

Other Known Limitations

• English only in this snapshot.
• Text + portrait art only, no audio, animations, or other media.
• Patch drift, see above.

Provenance

The provenance.json records:

• sts_game_version: the STS2 Steam version at fetch time
• source_fetched_at: UTC timestamp
• model and pipeline metadata

Citation

@dataset{sts2_cards_dataset,
  title = {Slay the Spire 2: Cards},
  author = {timothy22000},
  year = {2026},
  url = {https://huggingface.co/datasets/t22000t/slay-the-spire-2-cards},
  note = {Early Access snapshot; card data via nkhoit/spire-archive; game IP © Mega Crit}
}

Licensing

• Dataset: CC BY 4.0
• Pipeline code: MIT, see github.com/timothy22000/slaythespire-codex
• Game IP: Slay the Spire 2 is © Mega Crit.