Add arithmark-2.0-scratch-19m to the leaderboard (97.40%)
#4
by TobiasLogic - opened
Adds a from-scratch 18.9M parameter character-level model that scores 97.40% overall (99.28 / 99.60 / 89.40 across 1/2/3 operators), reproducible with benchmark_arithmark-2.0.py.
Model: https://huggingface.co/TobiasLogic/arithmark-2.0-scratch-19m
Training data (decontaminated against the 2500 benchmark items): https://huggingface.co/datasets/TobiasLogic/arithmark-2.0-train
Withdrawing this. On reflection this model is a narrow arithmetic specialist rather than a general model, and your Open SLM Leaderboard already classifies specialists separately. Closing to keep the baseline table clean. Thanks for the benchmark!
TobiasLogic changed pull request status to closed