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.
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
| Capability | Kate's implementation |
|---|---|
| Sequence understanding | New causal multi-head attention pilot alongside the older recurrent models. |
| Source recognition | Saved-source retrieval with exact excerpts, document identity and dates. Evidence is separate from model memory. |
| Autonomous actions | Model has no execution access. Explicit commands retain their bounded tool handlers. |
| Voice, vision, coding, translation | Separate 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.