MMLA-Orbit-1B

Added MMLA weights and the semantic adapter in this
release are distributed under GNU Affero General Public License version 3
(AGPL-3.0), as selected by the project owner. LICENSE contains the license text.

The model substrate and tokenizer files originate from openbmb/MiniCPM5-1B-SFT,
revision a60b37f1fc409c54e1e337b0723aaac6f92dfec0. Its upstream model card declares
Apache-2.0. Those files remain under Apache-2.0; the AGPL declaration does not
relicense them. The complete Apache-2.0 text is in licenses/Apache-2.0.txt and
the original model card is retained as licenses/UPSTREAM_README.md.
Upstream: https://huggingface.co/openbmb/MiniCPM5-1B-SFT

Original tensor values and tokenizer bytes are unchanged. The base configuration
is retained inside config.json, and model tensors are consolidated into
model.safetensors under separate namespaces.
The added runtime attaches omega without merging or overwriting the original
base tensors. Added task, trained-proposal and original-proposal MMLA tensors
are provided as safetensors.

results/PROVENANCE.json retains the original substrate, runtime, result and review
identities as historical provenance. config.json and MANIFEST.json bind this
consolidated package and its matching runtime.
The public API supports saving and restoring proposal-process state. Original
training optimizer states and training RNG snapshots remain in the internal
reproducibility archive, alongside restricted task records and raw evaluation
traces; they are not included in this public inference package. PyTorch,
Transformers, Safetensors and other dependencies retain their respective
licenses and are not redistributed as binaries in this release.