Livnium Labs · Model collection
Livnium's vector-collapse program has produced working systems for semantic embeddings, compact text generation, and natural-language inference. Each model has a real checkpoint, a runnable path, and a measured result.
The achieved models
These are the strongest trained results in the Livnium repository. They are presented as completed engineering achievements with their evaluation boundaries kept visible—not as abandoned experiments.
A pure vector-collapse model that learns lexical meaning from raw Wikipedia context—without an MLP, attention, a transformer, or pretrained embeddings.
A compact conditional language model that turns a hypothesis and NLI label into a contextual premise using learned semantic attractors.
A 3-way natural-language-inference classifier whose learned geometry moves sentence-pair representations toward entailment, neutral, and contradiction attractors.
What was achieved
The work is not one all-or-nothing claim. It is a sequence of concrete results: semantic geometry, fast conditional generation, and learned-attractor classification above strong simple baselines.
Noun Collapse learned semantic neighborhoods with one well per word and no separate neural output head, reaching 0.3616 on the noun subset of SimLex-999.
The 5.98M-parameter Premise Generator produced contextual short replies in roughly 6 ms median on CPU.
At 68.92% SNLI accuracy, the model beat the 61.5% hypothesis-only artifact by 7.42 points and the 60.7% GloVe-average baseline by 8.22 points.
Benchmark context
The headline results stand on their own. The remaining controls define the next research questions without erasing what the trained checkpoints already demonstrate.
Beyond the model collection
Active work extends the same emphasis on inspectable mechanisms and reproducible measurements into language, vision, dynamics, and combinatorics.
A character-to-word-to-conversation ladder exploring longer contextual reading and response generation.
Inspect the program →An investigation of attractor dynamics over image data, with every pixel represented as an attractor.
Inspect the program →Solvers, independently checked witnesses, and group-action constructions with reproducible verification.
Inspect the program →Explore the work
The code, model cards, evaluation records, and published Noun Collapse checkpoint are open for builders and research collaborators.