nikhilchandak/OpenForecaster-8B
OpenForecaster-8B by nikhilchandak is an 8 billion parameter language model, post-trained from Qwen3-8B, specifically designed for open-ended forecasting and predicting future events. It leverages reinforcement learning on the OpenForesight dataset and features a 32768 token context length. This model excels at providing calibrated confidence estimates, reasoning about uncertainty, and utilizing retrieved information to improve predictions, outperforming much larger models on forecasting benchmarks like FutureX for non-numeric questions.
Loading preview...
OpenForecaster-8B: Specialized for Future Event Forecasting
OpenForecaster-8B, developed by nikhilchandak, is an 8 billion parameter language model post-trained from Qwen3-8B. Its core purpose is to provide calibrated predictions for open-ended questions about future events. This model was fine-tuned using reinforcement learning on the extensive OpenForesight dataset, which comprises over 52,000 forecasting questions derived from global news.
Key Capabilities
- Calibrated Confidence Estimates: Provides reliable probability estimates when explicitly prompted.
- Uncertainty Reasoning: Capable of reasoning about various future scenarios and inherent uncertainties.
- Contextual Prediction: Effectively leverages retrieved information provided in the context to enhance prediction accuracy.
- High Accuracy: Achieves superior accuracy on forecasting benchmarks like FutureX for non-numeric questions, outperforming models significantly larger than 100B parameters (based on a July-August 2025 evaluation with a fair knowledge cutoff).
Good For
- Open-Ended Forecasting: Ideal for predicting outcomes of future events with nuanced reasoning.
- Decision Support: Useful in applications requiring probabilistic assessments of future scenarios.
- Research in AI Forecasting: A strong baseline for further research into scalable open-ended reasoning and predictive AI.
Note: The model's knowledge cutoff is approximately April 2025. For questions concerning events post-April 2025, it is recommended to use Retrieval-Augmented Generation (RAG) with up-to-date context.