Fine-tuning
Results from fine-tuning runs on Lambda GPUs, read-only.
Lab results
Muse Muse fine-tunes Meta's open models on GPUs: Spark 1.3 Contributor writes a training set, Muse Glimmer 30B learns it, and Spark 1.3 Contributor judges base against tuned on held-out prompts. Muse Spark 1.3 wrote and judged the runs before Oct 29, 2026, and each result's page names the models that did its work. Decision models are scored instead of judged: Muse Glimmer 30B learns to pick one label from a person's labeled examples, and base and tuned are checked on rows held out before training, with no judge.
1. Write
Spark 1.3 Contributor writes a training set for a task and holds some prompts out.
2. Train
Muse Glimmer 30B learns it with a LoRA adapter on a Lambda GPU.
3. Judge
Spark 1.3 Contributor scores base and tuned Glimmer on the held-out prompts, without knowing which is which.
Results
| Job | Model | Tuned vs base | Average score | GPU | GPU time | Cost |
|---|---|---|---|---|---|---|
| Made-up shop inbox, decide real (D8)Muse Glimmer LoRA · decision model | Muse Glimmer 30B | Tuned right on 78 of 78 (100%)Base right on 78 (100%) | Macro F1 base 1.00 · tuned 1.00held-out rows, no judge | 1x A100 SXM4 (40 GB) | 15 min | $0.50 |
| Made-up shop inbox, decide score (D8)Muse Glimmer LoRA · decision model | Muse Glimmer 30B | Base right on 78 of 78 (100%)Score only, nothing trained | Macro F1 1.00held-out rows, no judge | 1x A100 SXM4 (40 GB) | 9 min | $0.30 |
| Spark teaches Glimmer: brief writer (short prompt)Muse Glimmer LoRA | Muse Glimmer 30B | Tuned won 40 of 40Base won 0, 0 ties | Base 1.9 · Tuned 8.9out of 10 | 1x A100 SXM4 (40 GB) | 42 min | $1.39 |
| Spark teaches Glimmer: brief writerMuse Glimmer LoRA | Muse Glimmer 30B | Tuned won 11 of 40Base won 28, 1 tie | Base 6.8 · Tuned 3.6out of 10 | 1x A100 SXM4 (40 GB) | 48 min | $1.59 |
Datasets
| Dataset | Examples | Cost (est.) | Created |
|---|---|---|---|
| a cozy bakery helperAnswer questions about a made-up bakery: hours and allergens, warmly and briefly. | 32 train · 8 eval | $0.006 | |
| a bakery's warm counter helpertune a model on made-up bakery questions so it answers warmly and briefly | 32 train · 8 eval | $0.006 | |
| Made-up shop inbox (D8 test)Which team should answer this made-up customer message? | 274 train · 38 dev · 78 test | $0.00 | |
| Muse Muse brief writerYou are Muse Muse's brief writer. A user describes a small app they want, in one or two s… | 360 train · 40 eval | $1.51 |
Anyone can start a run. Start a fine-tune with the start code. Your runs stay private, and they never go to Hugging Face.