The first fully open model to bring the ILP paradigm into RLVR. We wire a Prolog interpreter straight into the reward loop and execute logic programs to grade the model. 📝 Blog • 💻 Training Code • 📊 Eval Code • 🤗 SLR-Bench • 🤗 Olmo 3.1 7B Think Open RLVR recipes center on math and code, and logical reasoning gets left behind. OlmoLogic fixes that. Starting from Olmo-3-7B-Think-DPO, we wire the paradigm of Inductive Logic Programming into Olmo 3’s RLVR receipe. OlmoLogic is post-train from-scratch on 56×H100 for 6 days straight (3,350 optimization steps) broadly outperming Olmo-3 7B Think with large gains on logical reasoning. Held-out logic suite: LogiGLUE, KOR-Bench, bAbI-16, CLUTRR, FOLIO, ProntoQA, RuleBERT, and abductive reasoning. All numbers from a single reproducible OLMES pipeline. > 📖 Full ablations, training dynamics, and the compute-matched control (Olmo 3.1 7B Think) are in the blog post. The model plays a train classifier: trains travel east- or westbound, each made of cars with properties (color, length, wall type). Given positive and negative examples plus background facts, it must write the shortest Prolog rule that…
Модальности:
Генерация текста
Области применения:
Логика и рассуждение Диалог / чат
Задача: Генерация текста
Автор: AIML-TUDA
Теги: olmo3, reasoning, logical-reasoning, rlvr, grpo, inductive-logic-programming, prolog, olmo
Лайков: 4 | Загрузок: 337
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.
