Post
56
Meta released Muse Glimmer 30B on Aug 10. We fine-tuned it the next day.
Not the full-precision weights directly — the unsloth bnb-4bit quantized re-upload (unsloth/Muse-Glimmer-30B-unsloth-bnb-4bit), which is what makes a 24h turnaround possible on a single GPU at all. Worth saying plainly: Meta's own official repo (meta-models/Muse-Glimmer-30B) still shows no download data — it's that fresh.
What we tuned it on: not new facts, a pattern. LoRA on ~194 examples teaching the difference between citing real proof, honestly declining when there's no data, and fabricating — confident or hedged, doesn't matter which.
Results on 20 held-out claims never seen in training:
- base model: 0/20
- tuned: 20/20
Training: 472.5s, loss 0.799 → 0.086.
Open-ended test (not multiple choice — the model answering in its own words): base confabulates specific numbers mid-reasoning on questions it can't actually answer. Tuned: declines cleanly, every time.
Dataset: SoulInPsyAbstract/specialist-cd-binary-honesty
Adapter: SoulInPsyAbstract/specialist-cd-muse-glimmer-lora
Meta's release: https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model
Same non-fabrication pattern also holds on Hermes-3-8B and Qwen2.5-7B, tested with the identical held-out set. Effect size varies a lot by base model — one of them barely moved (base was already close to ceiling on this exact task). More on that soon.
Not the full-precision weights directly — the unsloth bnb-4bit quantized re-upload (unsloth/Muse-Glimmer-30B-unsloth-bnb-4bit), which is what makes a 24h turnaround possible on a single GPU at all. Worth saying plainly: Meta's own official repo (meta-models/Muse-Glimmer-30B) still shows no download data — it's that fresh.
What we tuned it on: not new facts, a pattern. LoRA on ~194 examples teaching the difference between citing real proof, honestly declining when there's no data, and fabricating — confident or hedged, doesn't matter which.
Results on 20 held-out claims never seen in training:
- base model: 0/20
- tuned: 20/20
Training: 472.5s, loss 0.799 → 0.086.
Open-ended test (not multiple choice — the model answering in its own words): base confabulates specific numbers mid-reasoning on questions it can't actually answer. Tuned: declines cleanly, every time.
Dataset: SoulInPsyAbstract/specialist-cd-binary-honesty
Adapter: SoulInPsyAbstract/specialist-cd-muse-glimmer-lora
Meta's release: https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model
Same non-fabrication pattern also holds on Hermes-3-8B and Qwen2.5-7B, tested with the identical held-out set. Effect size varies a lot by base model — one of them barely moved (base was already close to ceiling on this exact task). More on that soon.
