SPROG-9M — a 9.37M parameter model trained from scratch to solve GSM8K-style math without using an LLM at inference.
The model, codelion/sprog-9m, predicts symbolic programs over number slots, then a deterministic executor does the arithmetic. With a simple verifier, it reaches ~11.8% on GSM8K test.
We also released the dataset: codelion/gsm8k-synth, 117K validated synthetic GSM8K-style problems.
Tiny model, no pretraining, no LLM at inference, runs on a laptop.
In algebraic geometry, embedding tropical varieties into toric varieties relies on the process of tropicalization and the use of toroidal embeddings. Rather than embedding as geometric spaces, tropical varieties are embedded as polyhedral fans or cone complexes within the fans of target toric varieties to study compactifications and intersection theory.
Toric Varieties: Geometric varieties constructed from combinatorial data known as fans (collections of polyhedral cones). Tropicalization: A procedure that maps an algebraic variety X (typically a subvariety of an algebraic torus T) to a polyhedral complex in a real vector space via valuations. Toric Closures: When you embed a tropical variety in a compact toric variety, the boundary points correspond to limits of points in the variety going towards the boundary of the torus. Embedding Methodologies...Toric Compactification: Any tropical variety (a rational polyhedral fan) can be compactified by embedding it into a complete toric variety. The fan of the toric variety is chosen to contain the tropical variety as a subcomplex, which encodes the combinatorics of the compactification. Toroidal Embeddings: For varieties that are not strict toric varieties, they can often be mapped into toric varieties using the theory of toroidal embeddings. This involves equipping the tropicalized spaces (cone complexes) with balancing conditions and intersection theories.
I have downloaded about 8.7 million of 2048x1024 (or 1664x832 for older captures) panoramas randomly distributed all over the world that I toyed around with training geo-guessing models. Is anybody interested in a dataset of Google Street View panorama images along with latitude-longitude coordinates?
To make revising LLM architectures and training methods faster, I created a deck of 180 visual flashcards. It started as a personal hobby, but slowly became cheat code for reviewing LLM concepts before technical interviews. People love it!
Inspired by the Nemotron Diffusion recipe, check out dhara-250m: a 250M experimental language model that supports three decoding modes from one set of weights: autoregressive, block-diffusion, and self-speculation.
It is small, easy to try, and meant for exploring diffusion-style decoding and latency tradeoffs in compact LMs.