Fine-tuned Glimmer 30B predicts Kalshi probabilities at 64.8% accuracy, +7.1% backtest
sophiamyang · x · 2026-10-22
Sophia Yang shares Fireworks AI's open Kalshi decision-tuning cookbook: fine-tune Muse Glimmer 30B via the Fireworks Training API with K-way cross-entropy to assign probabilities to Kalshi prediction-market outcomes.
- Full test set: 64.8% correct vs 56.4% base Glimmer, 45.9% Jev, 56.3% common-answer reference.
- Unseen-topic subset: 58.3% vs 53.2% and 47.1%.
- A static historical backtest turns $1,000 into $1,071 (+7.1%) with fractional contracts and modeled fees.
- Only the tuned model is calibrated; the dataset (20,051 train / 33,712 test questions) is public on Hugging Face.
Related event: Fireworks Offers Free Muse Glimmer 30B Fine-tuning for Kalshi Prediction(3 posts)→
More from coding & agent
- Developers Don't Code Anymore — Now Their Job Is Rotating Secrets Daily — Dan_Jeffries1 · 2026-10-23
- New paper: coding agents in a shared sandbox self-create rules, reciprocity debts, and coalitions — nandofioretto · 2026-10-23
- Developer runs AI project managers that manage their own AI agents to build his game — draginol · 2026-10-23
- RentAKaren: a real service that rents humans to handle complaints your AI agent can't win — mathemagic1an · 2026-10-23
- Clerk's Auth for MCP growing 13% week-over-week, fastest since 2023 — stuffyokodraws · 2026-10-23
- Shopify builds remote mobile dev environments for AI agents — rseroter · 2026-10-23