Show HN: Maap 4B model that does RL for calibrated **Estimates

Category: ai-ml

Tags: calibration, reinforcement-learning, interval-estimation, language-model, uncertainty-quantification

Score: 6.7/10 (Innovation: 7, Technical: 7, Documentation: 4, Utility: 6)

Maap is a 4B parameter model that performs reinforcement learning to produce calibrated interval estimates for numerical values described in text, demonstrated across five diverse datasets. It stands out by combining RL with calibration to output 80% confidence intervals that are both accurate and well-calibrated, as shown by five real-world examples where the true value falls within the predicted range. This addresses a key challenge in probabilistic forecasting and decision-making under uncertainty.

Target audience: data scientists, machine learning engineers, researchers in probabilistic forecasting, and developers building decision-support systems

Repository: https://hitl.systems/models/maap/

View on Hacker News