Add TextModel-v1 to leaderboard

#51

TextModel-v1 (bench-labs), 122.7M params trained from scratch (85% FineWeb-Edu / 15% synthetic causal-reasoning data, 1.64B tokens). GCI 55.044 on the full 5000-item test set, BLiMP 80.31, ARC-Easy 49.71, WikiText-2 byte-ppl 2.628. Model card and eval methodology: https://huggingface.co/TobiasLogic/TextModel-v1

CompactAI changed pull request status to merged
GlintResearch org

Nice GCI! You just beat the top GCI scorer with 3M less parameters(they had 125M)!
Well done!

This comment has been hidden (marked as Off-Topic)

Oh also tobias, jsyk this leaderboard runs of acc (non norm) so make sure you cite that

@Datdanboi25 will do, I'm not currently carrying my laptop right now, will add that for sure

Sweet, usually adds a few points to the arc score

Sign up or log in to comment