← Open K.A.T.E. · Earlier recurrent experiments
ORIGINAL CODE · OWNER-TRAINED WEIGHTS

Kate's new model foundation

An original causal-attention decoder, byte tokenizer and training pipeline. Public research informed the design; implementation code and pretrained weights were not imported.

This is a research pilot, not a finished intelligent assistant. It stays separate from live chat and has no tool access. Main chat can use the separately approved local Qwen model and existing explicit command handlers. This research checkpoint is not the conversation model.

Model parameters—Random initialization
Development exact answers—Inspect outputs; not a broad benchmark
Assistant promotionOffNo action execution

Training result

Loading local result…

The tokenizer sees only training examples. Validation chooses the checkpoint. Test examples are measured afterward. All 34 examples are original synthetic development material; they do not establish general understanding.

What the research changes

CapabilityKate's implementation
Sequence understandingNew causal multi-head attention pilot alongside the older recurrent models.
Source recognitionSaved-source retrieval with exact excerpts, document identity and dates. Evidence is separate from model memory.
Autonomous actionsModel has no execution access. Explicit commands retain their bounded tool handlers.
Voice, vision, coding, translationSeparate data, model and evaluation tracks; the new text pilot does not provide these abilities.

Inspect experimental generation

Expect incomplete or incoherent output. Sign in to Kate in this browser first.