From the source
From the source
Study uses Ai2's open stack to trace social reasoning origins in Olmo 3.
Georgia Tech researchers used Ai2's open Olmo ecosystem—including Olmo 3, Dolma 3, WebOrganizer, OlmoEval, and OLMES—to trace the training data origins of social reasoning in a language model.
They found that dialogue-rich, interpersonal writing categories most influenced Olmo 3's social reasoning, and a causal test showed that removing the most influential literature documents significantly reduced the model's SocialIQA score.
From the source
Matlin and Chakraborty then ran a causal test: if they deleted the data that most shaped an ability, would the model get worse at it? They had Olmo 3 "unlearn" the most influential documents in Dolma's literature category, which was the category most tied to Olmo 3’s social reasoning. Its SocialIQA score dropped further than when they removed random documents from that category.
allenai.org
