AbstractPhil/diffusion-pretrain-set-ft1-1024
Viewer • Updated • 1.14M • 92
AbstractPhila PRO
AbstractPhil
AI & ML interests
datasets, research papers, experimentation, vision, classification, text encoders, tokenization, llms, diffusion, distillation, and more.
Recent Activity
posted an update about 2 hours ago
It's still just a whisper now as it trains the mathematics directly into a testable and representable arc.
The first aleph geometric vocabulary LLM is being trained using the same trigram aleph-void paradigm based exactly on this math and curve. The curve is testing the plateau, the structure, the underlying behavior, and the responses of the model to the actual formatting and relational systems.
She handles a vocabulary of 12.8m potential grams in vocabulary form. This prototype LM model predicts multiple next tokens simultaneously, not just one - many. Preliminary sequence length of around 1024 and context window of 4096 operating on K64.
Currently, this math is directly testing the upper bounds of this scaling curve. It must find the plateau for a couple benchmarks before we can ascend with large spectrum trained vocabulary data.
Furthermore, the context window when correctly routed with cantor fractal routing is mathematically capable of massive... massive attention routes with massive sequences without OOM, as per the scaling theorem and the applied stability. That when paired with aleph routed attention we have a solver that can not only solve, but it WILL solve as a guarantee. This is the same principality as the alephs, geometric structure, scaling principles, and the experimental line from the geofractal runs.
Everything is snapping together like legos with testable structural bounds for each element built from battle-tested analysis from hundreds of experiments. The theorems are structurally valid, the models are structurally valid, the system is coalescing, and the math lines up.
She cooks the first field scaling test now for big numbers
https://huggingface.co/AbstractPhil/geolip-aleph-void/tree/main/experiments/exp_007_aleph_routed_attention
Trigrams are the language this model knows, and it's nowhere NEAR the only language the model can learn. The language can be anything, of any organization, of any type when formatted into the aleph format. updated a model about 17 hours ago
AbstractPhil/geolip-aleph-void updated a dataset about 19 hours ago
AbstractPhil/diffusion-pretrain-set-ft1-1024Organizations
