DSEG
DSEG-Char — a character-level Chinese poetry model in which every character is a node in a 5,000-node graph, and writing a line is a walk over that graph.
Overview
DSEG-Char is a character-level generator in which every character is a node in a 5,000-node graph, and writing a line is a walk over that graph. The three terms that select the next character — state, edge and context — are kept separate, so each step can be asked which one carried it; a browser visualiser colours the walk accordingly. The trained weights ship with the code, so it runs on a fresh clone.
What it does
Continues a line
Give it a prefix and it writes on in classical Chinese verse, decoding by top-k, nucleus, temperature or argmax. Trained on 24,000 Tang poems over a 5,000-character vocabulary.
Explains every step
Because state, edge and context are separate terms in the score, each character comes with the one that actually pushed it up — attribution per step, not a post-hoc probe.
Shows the walk
A browser visualiser draws the 5,000 nodes, lights the path as it is written, and colours each hop by the term that carried it.
夜半聞鐘聲 → ,孤雲入天閣。清晨出松蘿,月上山中宿。 秋水連天碧 → ,天涯一月長。不辭千里曲,應逐白雲時。 江上一孤舟 → ,山中一片雲。何年此地遠,萬里一秋深。
The corpus is in traditional characters, so prefixes are too. Nothing else is needed: the weights are in the repository, so a fresh clone generates without downloading a corpus or training anything.