My Name is Riemann Prerelease - Basic Demo
13.1B parameters. Grown rung by rung from OpenPeerLLM and fine-tuned with the OpenPeer-AI-NTK-Trainer gate controller instead of LoRA. Weights: OpenPeerAI/My-Name-is-Riemann-Prerelease.
This is a prerelease base model, trained on about 2.2 billion tokens of encyclopedic text, agentic coding transcripts and source code. It is not instruction-tuned and does not chat. The tabs below show what it has actually learned: completing text in the register it was trained on, scoring how expected each token of a passage is, and, most tellingly, telling fluent language from noise.
Give a subject and an opening clause in encyclopedia register; the model continues it. Two example articles are prepended as context.
| Subject | Opening clause |
|---|
Paste any text. Each token is coloured by how surprised the model was to see it. Fluent text in a register the model knows comes out green; errors, nonsense and unfamiliar registers come out red. With the controller enabled you also see how much the gates change the model's view of the text.
Six fixed strings, from fluent encyclopedic English down to random words, scored by the model. A working language model separates these by orders of magnitude.
Licensed under the Open Innovation License v2 (OPNL-2) and MIT. Developed by OpenPeer AI, Riemann Computing Inc., and Andrew Magdy Kamal.