YarrowLab/yarrow-forecast

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 10, 2026License:cc-by-nc-4.0Architecture:Transformer Open Weights Featherless Exclusive Cold

YarrowLab/yarrow-forecast is a 27 billion parameter model designed specifically for forecasting U.S. party primary and nomination outcomes. Unlike traditional LLMs, it provides a calibrated probability between 0 and 1 in a single forward pass, without generating text or abstaining. This specialized architecture allows for deterministic, fast inference (0.12–1.6 seconds per question) and operates locally without external API dependencies, making it ideal for high-throughput, cost-efficient probabilistic forecasting in its validated domain.

Loading preview...

Overview

YarrowLab/yarrow-forecast is a specialized 27 billion parameter model engineered to predict outcomes for U.S. party primary and nomination elections. It uniquely provides a calibrated probability (between 0 and 1) for an event resolving "YES" in a single, deterministic forward pass, eliminating text generation and abstention. This design ensures rapid inference, with batch processing times ranging from 0.12 to 1.6 seconds per question.

Key Capabilities

  • Direct Probability Output: Returns a precise probability from the next-token distribution, rather than generating explanatory text.
  • Local & Cost-Efficient Inference: Operates entirely on your own GPU, incurring no per-call fees or reliance on third-party APIs, making marginal costs near zero.
  • High Throughput: Achieves fast, deterministic scoring without sampling or token-by-token decoding.
  • Specialized Domain Expertise: Validated specifically for U.S. party primary and nomination elections, demonstrating strong performance within this niche.
  • Further Trainable: Supports additional training via LoRA updates, maintaining calibration accuracy.

Performance Highlights

On a set of 150 U.S. party primary and nomination questions, yarrow-forecast achieved a Brier score of 0.175 and a directional accuracy of 76.0%. This performance surpasses Claude Sonnet 4.6 (70.1%) and GPT-5.5 (65.5%) in directional accuracy, and outperforms GPT-5.5 (0.185), GPT-5.4-mini (0.242), and OpenForecaster-8B (0.265) in Brier score. The model consistently answered all 150 evaluation items with 0% abstention.

Good for

  • Developers requiring highly specific, probabilistic forecasts for U.S. political primary and nomination outcomes.
  • Applications needing fast, deterministic, and cost-effective inference without external API dependencies.
  • Scenarios where a direct probability output is preferred over generated text explanations.
  • Researchers and analysts focused on political forecasting who can leverage its specialized domain validation and further fine-tuning capabilities.

Finetunes
1 models