GPokeT2-dataset / README.md
iamthinbaker's picture
Update dataset card
65e40f3 verified
|
Raw
History Blame Contribute Delete
28.6 kB
metadata
license: apache-2.0
tags:
  - pokemon
  - sprites
  - pixel-art
  - gpt2
  - ascii-encoding
pretty_name: GPokeT2  Pokémon Sprite Dataset
size_categories:
  - 10K<n<100K
task_categories:
  - text-generation

🎮 GPokeT2 — Pokémon Sprite Dataset

Pokémon sprites from all mainline Gen 3 and Gen 4 games, encoded as ASCII token sequences and paired with rich metadata conditioning labels. Used to train GPokeT2.

Pokemon sprite ASCII representation Train the model
-> -> GPT2-Small

📦 Source data

All sprites were sourced from Veekun, a community sprite repository. Metadata (types, generations, evolution chains) was gathered via PokéAPI.

Generation Game Sprites
Gen 3 Pokémon Emerald 1 600
Gen 3 Pokémon FireRed / LeafGreen 312
Gen 3 Pokémon Ruby / Sapphire 837
Gen 4 Pokémon Diamond / Pearl 2 528
Gen 4 Pokémon Platinum 2 556
Gen 4 Pokémon HeartGold / SoulSilver 2 560
Total 10 393

🔤 Pixel → ASCII encoding

Each 64×64 sprite is serialized as a flat sequence of ASCII characters. Each pixel is quantized to 4 levels per channel (R, G, B ∈ {0, 1, 2, 3}) and packed into a single character:

char = chr(R×16 + G×4 + B + 59)   # 64 possible color chars
char = '~'                          # white / transparent pixel

This produces a vocabulary of 65 pixel tokens, plus 64 special row-marker tokens ([ROW_00][ROW_63]) that delimit each row. A full sprite is therefore a sequence of 64 rows × 64 pixels = 4 096 tokens.

The encoder lives in src/application/slv/pokedex/encoder/pokemon_encoder.py (PokemonEncoder).


🔁 Data augmentation

Each sprite is augmented to produce 12 variants:

Technique Factor Description
Horizontal flip ×2 Pixel order reversed per row at the ASCII level
Color shift ×6 All 5 non-identity RGB channel permutations + original palette

Both augmentations are independent and combined: 1 sprite → 2 flip variants × 6 color variants = 12 samples.
Final training set: ~124 700 sequences.


🏷️ Conditioning labels

Each sample carries the following metadata fields:

Field Values Description
type1 0–18 Primary type (18 types + unknown)
type2 0–19 Secondary type (18 types + none + unknown)
generation 0–9 Game generation (Gen I–IX + margin)
evo_stage 0–3 Basic / Stage 1 / Stage 2 / other
has_evolution 0–1 Whether the Pokémon can still evolve
is_shiny 0–1 Normal vs. shiny palette
color_shift 0–5 Which RGB permutation was applied
pokemon_idx 0–N Unique Pokémon identity index

Available types (index order):

0 ⬜ normal 6 🥊 fighting 12 🔮 psychic
1 🔥 fire 7 ☠️ poison 13 🐛 bug
2 💧 water 8 🌍 ground 14 🪨 rock
3 ⚡ electric 9 🌪️ flying 15 👻 ghost
4 🌿 grass 10 🐉 dragon 16 🌑 dark
5 🧊 ice 11 ⚙️ steel 17 🧚 fairy

🙏 Acknowledgements

Inspired by matthewRayfield/pokemon-gpt-2.

  • PokéAPI — metadata (types, generations, evolution chains)
  • Veekun — original 64×64 PNG sprites