ml-pokedex — ONNX weights

Byte-level BPE decoder-only transformers trained on RDF (not English), over the Pokémon Knowledge Graph. These are the weights behind the in-browser demos at github.com/alganet/ml-pokedex; each file is a KV-cached ONNX graph (O(T) generation) that runs client-side with ONNX Runtime Web.

file	params	size	powers
memoriser.onnx	~19.8M	~80 MB	The Memoriser — recalls a real Pokémon's stored facts
combo.onnx	2.86M	~11 MB	The Minter (invents a creature) and the Reasoner (derives weaknesses)

Both share the tokenizer in the app repo (assets/bpe.json, byte-level BPE, vocab 1024) and consume/produce Turtle. Per-model dims are in the app repo's models/*.json.

Inputs / outputs

The graph takes ids (1,T) plus a past key/value cache past_k/past_v (L,1,H,P,D) and returns logits (1,V) at the last position plus the grown cache. Prompt framing:

• memoriser: [BOS] + encode(prompt) + [SEP], then decode the answer (greedy; an optional IRI trie in the app forbids hallucinated identifiers).
• combo: [BOS] + encode(prompt), continue the document (sampled for the Minter, greedy for the Reasoner).

Honest note

Judged as a database the memoriser loses to gzip (the graph gzips to 0.17 MB; the model is ~80 MB at worse recall). The interesting behaviour is what the 11 MB combo model does on a fresh IRI it cannot look up: invent a coherent creature and derive its weaknesses from the type chart. See the app's about pages for the measured results.

License / attribution

Trained on data derived from the Pokémon Knowledge Graph. Pokémon and all related names are © Nintendo / Game Freak / The Pokémon Company. Non-commercial research artifact, not affiliated with or endorsed by them.