Fine-tuning Meta's Muse Glimmer on Kalshi Prediction Outcomes
Melvin Vivas · X post · 2026-10-22 · Open on X
Topics: Fine-tuning & Model Customization · Level: intermediate
Summary
A quoted post says a team fine-tuned Meta's Muse Glimmer model on Kalshi prediction-market outcomes using supervised fine-tuning (SFT). They report that one training pass beat a baseline in prediction accuracy and in a hypothetical backtest. The creator wonders what the fine-tuning would cost. The quoted post says SFT training is free until Nov 30.
Key points
- Muse Glimmer (Meta) was fine-tuned with SFT on Kalshi prediction-market outcomes.
- One training pass reportedly put it 11.2 percentage points ahead of the 'Jev' baseline in prediction accuracy.
- In a backtest it reportedly returned +7.1% against the baseline's −8.1%, using hypothetical money only.
- SFT training is offered free through Nov 30, and a training recipe is mentioned, but its link is not included in the post.
- Real-world outcome data (like prediction-market results) can serve as labels for fine-tuning a forecasting model.
Resources mentioned
- Muse Glimmer · tool · research.meta.ai · free
A Meta model that the quoted post fine-tuned with SFT for prediction tasks. - Kalshi · website · kalshi.com · check price · open in a browser to verify
A prediction market whose event outcomes were used as training labels.
Try this
- Try the free SFT fine-tuning offer before Nov 30 (the recipe link is not in the post).
- Fine-tune an open model on resolved prediction-market outcomes and backtest its forecasts against a baseline.