Story

Hey HN! I RL-tuned an open-source LLM (gpt-oss-120b — 120B MoE, but only 5.1B active params) to predict "What will Trump do?" in any situation, trained on nothing but public news collected automatically from search queries. The trained model beats GPT-5, and both dataset and trained model are open sourced.

Data generation: Generated 2,108 binary forecasting questions from just a search query and a date range using the Lightning Rod SDK (https://github.com/lightning-rod-labs/lightningrod-python-sd...). Questions are generated from historic news articles — like "Will Trump impose 25% tariffs on Mexico by March 1?" — and resolved by checking what actually happened after the deadline. No human annotation — the whole pipeline is automated.

Training: GRPO with Brier score as the reward signal. LoRA rank 32, 50 training steps.

Results: Slight accuracy edge over GPT-5 (Brier 0.194 vs 0.200), but big gains in calibration — the RL-tuned model produces much better probabilities (ECE 0.079 vs 0.091).

Dataset: https://huggingface.co/datasets/LightningRodLabs/WWTD-2025

This is a fully automated way to spin up domain expert LLMs from public web data with just a few search queries, no labeling/annotation required.

I’d love any feedback, or suggestions for what domain expert to train next!