Skip to content
News
Research
All news

Research · ORCNEIT

Introducing Research on ORCNEIT Lab

Our first two papers examine actual training of our language model: a base-training run and dialogue fine-tuning. Each includes measurements, charts, methods and limitations.

Published:

From plans to measurements

The ORCNEITGPT page describes the system we want to build. The new section answers a different question: what have we tested, and what did we observe? Neither account can substitute for the other.

Research presents each study separately: its question, test conditions, measurements, findings and limitations. Experiment dates and website publication dates are shown separately. Chart values are available in tables and as CSV downloads.

First paper: base-model training

We report a run conducted on 22–23 August 2026. The model started from a fresh random initialization and completed 80,000 steps. It processed 327.68 million training tokens, including repeated exposure, not that amount of unique text.

The paper shows training and development loss alongside a separate capability evaluation. Training completed, but the model failed the combined quality gate. That finding belongs beside the learning curve: lower prediction loss does not establish a ready assistant.

Second paper: the effects of dialogue fine-tuning

Our second study compares the base and fine-tuned models on identical tasks. Fine-tuning improved termination and reduced severe degeneration, but too few tasks were solved correctly for assistant acceptance.

This is not a finished-product announcement. It is a specific experiment involving weight updates, measured improvements and unresolved problems.

Two papers with methods, charts and result tables.

Explore Research