OfflineBase/Muse-Glimmer-30B-Moss-Test-LoRA

Muse Glimmer 30B Moss Test LoRA (job 1898e46c)

Built with Muse Spark Open. A community fine-tune, not an official Meta release.

A short test tune on an invented hobby. Never a product-quality tune. This repo is private. Nothing here has had the license and personal-data review that a public release needs.

How it was made

LoRA r=16 on OfflineBase/Glimmer-30B-2026-Knowledge-unsloth-bnb-4bit at 6f12bf52, 120 steps, 900 training rows, H100 SXM5. Loss 1.37 at the end, peak 22.44 GB. Adapter sha256 starts 927f00f6.

Data

Made-up. The Zubrelian moss guide is an invented hobby, from a fictional planet. Written through Meta's Model API with Muse Spark 1.3 Contributor, under Meta's terms for that API.

Measured

40 kept-aside prompts, judged blind by Muse Wonder: tuned better on 40 of 40, 7.93 against 2.5. An easy test on a short run. All 40 judge notes credit the format or invented words, so it learned shape and vocabulary, not facts.

Not run on a Jetson in any log we have. Measured on an x86 stand-in under a memory limit, which isn't a Jetson.

License

The base is a Derivative of Muse Spark Open, so this adapter is too. Built with Muse Spark Open. Keep LICENSE, MUSE_MODEL_LICENSE.md, USAGE_POLICY.md and NOTICE with every copy. Both usage policies say the models are not meant for people under 18 and bar presenting outputs as written by a person. Not an official Meta release. No warranty.