COLM Paper Traces OLMo3-7B Capabilities to Dolma3 Data
Glenn Matlin announces COLM 2026 paper using gradient attribution on Dolma3.
Glenn Matlin posted that a COLM 2026 paper traces capability provenance of OLMo3-7B across Dolma3 with gradient-based training data attribution. A visible reply notes signed influence heatmaps across 576 bins that show SocialIQA spreading positive mass broadly. Matlin supplied links to the paper and code in the post. The thread supplies the first-party announcement of the researchers' own work; no independent corroboration of results appears in the packet.
Combined views
11.6K
2 posts, first seen 10h ago
Reactions
2 postsGM 
Glenn Matlin@GlennMatlin🤔Where in trillions of pre-training tokens do capabilities actually come from? 💥In our new COLM 2026 paper, we trace the capability provenance of OLMo3-7B across Dolma3 using gradient-based training data attribution 🧵 Paper: Code: https://github.com/HCAI-Lab-GT/capabilibara…
- likes: 192
- replies: 11
- bookmarks: 161
- reposts: 22
GM Glenn Matlin@GlennMatlinSigned influence heatmaps across 576 bins (24 topics × 24 formats): SocialIQA (social reasoning) spreads positive mass broadly across topics. The other three benchmarks concentrate positive mass heavily in Documentation and Structured Data.
- likes: 7
- replies: 0
- bookmarks: 2
- reposts: 2
