§ Corpora & models
653,530 rows, one corpus.
The Aureth Corpus, drawn to scale: every dot is 64 rows. Sort it by the behaviour a row trains, by the PMI dimension it is labelled with, or by how it trains: a supervised example or a preference pair.
Generated with NousResearch/Hermes-4.3-36B from 125 prompts · half supervised examples, half preference pairs · quality score 0.65–0.95, mean 0.88 · apache-2.0 · counts from the Hugging Face dataset viewer, 2026.11.20 · open the corpus ↗
I Every public corpus
14 datasets, all open.
Row counts and sizes as the Hugging Face dataset viewer reported them on 2026.11.20. Where the viewer has not indexed a dataset, we say so rather than guess. Bars are on a log scale.
5 Aureth
-
Aureth-Corpus-Hermes4.3-Generated
OUSIA PMI-Aligned DPO Corpus
653,530 rows
-
Aureth-SFT-Curriculum
236,060 rows
-
Aureth-DPO-Curriculum
3,765 rows
-
Aureth-Agent-SFT-Robust
243,291 rows
-
Aureth-V3-Training-Data
rows not indexed by the viewer
5 Aurelia
-
aurelia-causal-events
Aurelia — Causal Event Stream
560,428 rows
-
aurelia-civilization-metrics
Aurelia — Civilization Metrics Trajectories
12,600 rows
-
aurelia-federation-causal
Aurelia — Federation Causal Graph (Events + Edges)
rows not indexed by the viewer
-
aurelia-npc-population
Aurelia — NPC Population Snapshot
rows not indexed by the viewer
-
aurelia-runtime-corpus-v2
Aurelia Runtime Corpus v2
2,500 rows
2 Vireth
-
Vireth-the-Heretic
rows not indexed by the viewer
-
Vireth-DPO-Preference-Pairs
rows not indexed by the viewer
2 Benchmarks & traces
II The models
The Aureth family.
- Aureth 9B “The Architect” on Qwen/Qwen3.5-9B-Instruct
- Aureth V2 · 4B on Qwen/Qwen3.5-4B-Instruct
- Aureth 4B on Qwen/Qwen3.5-4B
- Aureth 0.8B on Qwen/Qwen3.5-0.8B
Benchmarks for the 4B and 9B are on the Aureth dials.