| --- |
| base_model: meta-models/Muse-Glimmer-30B |
| base_model_relation: adapter |
| license: other |
| license_name: apache-2.0-and-muse-model-license |
| license_link: LICENSE |
| library_name: peft |
| pipeline_tag: text-generation |
| tags: |
| - knowledge-update |
| - built-with-muse-spark-open |
| - lora |
| --- |
| |
| # Muse Glimmer 30B Knowledge, Stage-B-R30 LoRA adapter |
|
|
| > **Built with Muse Spark Open.** A community fine-tune, not an official Meta release. |
|
|
| LoRA adapter from the Glimmer 2026 knowledge run: Stage B on the r30 mix, the run before r30b (the released Knowledge model). This repo is private. Nothing here has had the license and personal-data review that a public release needs. |
|
|
| ## How it was made |
|
|
| 60,000 rows, 18,000 with reasoning, from the stage 2 adapter. The model trained on question-and-answer pairs that Muse Spark Open and Muse Glimmer wrote about passages of 2026 public text. Not made-up data. Adapters load on Meta's BF16 base. |
|
|
| The `adapter_config.json` base path was changed from a local machine path to `meta-models/Muse-Glimmer-30B`. Nothing else was changed. |
|
|
| ## Measured |
|
|
| Q4_K_M 26.2% of 2,500, 40 of 40 old facts. See the table in the GGUF repo. |
|
|
| Greedy decoding, on the run log's test sets. Not run on a Jetson in any log we have. |
|
|
| ## Known limits |
|
|
| These come from tests of the Knowledge family, and not all were repeated on this file. |
|
|
| - It makes things up and states wrong answers as fact. Don't rely on it for facts without a source. |
| - The training data comes from public news and web text, including publishers' pages collected through CC-NEWS. Publishing weights trained on it is an open question in our run log. |
| - This model is not the released Knowledge model. That one is `OfflineBase/Muse-Glimmer-30B-2026-Knowledge`. |
|
|
| ## License |
|
|
| Fine-tune of Muse Glimmer 30B (`meta-models/Muse-Glimmer-30B`, revision `cec94fcfd6512f2cfe708138c32533b346ea4f69`), which Meta releases under the Apache License 2.0. |
|
|
| **Built with Muse Spark Open.** Part of the training data was written by Muse Spark Open, so this model is a Derivative under the Muse Model License. These terms are carried from `OfflineBase/Muse-Glimmer-30B-2026-Knowledge-LoRA`: |
|
|
| - Keep `LICENSE`, `MUSE_MODEL_LICENSE.md`, `USAGE_POLICY.md` and `NOTICE` with every copy you share. |
| - Show "Built with Muse Spark Open" in your product's interface, documentation, website or settings menu. |
| - Follow the Muse Glimmer Usage Policy and the Muse Spark Open Usage Policy. Both say the models are not meant for people under 18, and both bar presenting outputs as written by a person. Tell your users the known dangers. |
| - Keep every other condition of the Muse Model License, including section 3. |
|
|
| Not an official Meta release. Meta has not reviewed or endorsed it. No warranty. |
|
|
