QIRA › LOLM — Latent Order Language Model Research Products Contact All of Qira ↗ Every day LOLM scans the Hugging Face dataset hub for the things it actually wants to be good at — reasoning, tool-use, retrieval, verification — filters each candidate by license, quality, duplication and poison, and writes a source-backed memory it can reuse immediately. What it does not do is quietly retrain on random data forever. Weight changes are gated behind an eval. The receipt below proves exactly what changed — and what didn't.
Each relevant, clean dataset becomes a structured memory — repo, license, quality, topic, with the Hugging Face URL as its source. LOLM is smarter the same day, with a citation.
Those memories are queryable, and license-clean, high-signal datasets are queued as training candidates — an index of what could teach the controller, not raw downloads.
Adapter training runs only behind an eval wall: it must beat the current model, inside a regression budget, on a clean license trail. It never fires on autopilot.
Commercial-license, high-quality datasets that passed every filter. They wait here until an eval proves training on them helps. Queued count is not a weight change.
/api/demo/hf/dashboard. Dataset metadata is from the public Hugging Face Hub; memory is
source-backed; weight training is eval-gated and currently reports
model_weights_changed: false.