Add TextModel-v1 to leaderboard
#51
by TobiasLogic - opened
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
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
Sweet, usually adds a few points to the arc score