| Dohnuts |
| Copyright 2026 PsiACE and Dohnuts contributors. |
|
|
| The source code is licensed under Apache-2.0. Model weights and source datasets |
| are separate artifacts with their own terms. |
|
|
| Qwen3.5-9B is developed by the Qwen team at Alibaba Cloud and is licensed under |
| Apache-2.0. Jolt trains a language LoRA adapter and a decision head over this Qwen |
| base. The Qwen base weights are fetched from the original publisher. |
| https://huggingface.co/Qwen/Qwen3.5-9B |
|
|
| The RLCD objective and evaluation protocols follow Laya by Nandha Kishor M and |
| Laya Vision by Reed Richards. Both projects are licensed under Apache-2.0. |
| Pinned references and implementation details are recorded in docs/rlcd.md and |
| docs/upstream-alignment.md. |
| https://github.com/NandhaKishorM/laya |
| https://github.com/r33drichards/laya-vision |
|
|
| JevBench supplies the public benchmark tasks and scoring protocol. Published Jev |
| outcomes are reference measurements, not locally executed Jev predictions. |
| https://github.com/fstandhartinger/jevbench |
|
|
| Bub supplies the agent SDK. Flash Linear Attention supplies fused operators. |
| They are installed as dependencies, not bundled copies of their source code. |
| https://github.com/bubbuild/bub |
| https://github.com/fla-org/flash-linear-attention |
|
|
| Dataset sources, provenance, and terms are listed in docs/data-and-evaluation.md. |
| Raw datasets, source images, and third-party model weights are not distributed |
| with this source repository or Python package. Reference names do not imply |
| endorsement by their authors. |
|
|
