Use the custom game-control task consistently, preserve training-stage tags, correct the 256-frame inference cache, close verified replay audits, link the collection and released code, and retain Slippi-AI attribution. Model weights and inference source are unchanged.
29 days ago
Redesign model card with visual guides and evaluation insights
Add stage-specific artwork, benchmark charts, architecture diagrams, training lineage, tested Transformers examples, and an evaluation manifest with source locators and protocol context.
Add verified Transformers package and safetensors weights
Complete Faynt inference package. Exact tensor equality and native inference parity verified for the selected checkpoint. Includes configuration, policy and codec code, structured inputs, model card, provenance, and MIT license.
