Tiny-LM efficiency track (Glint leaderboard). Auto-refreshes every 20s · …
NOW TRAINING — the final
GoLLeM-v5 64M (14×576) and
128M (16×768): 760,000 steps / 24.9B tokens each, the flagship recipe on ARC-MIX with test/validation overlaps and documents marked “CC BY-NC-SA” removed.
The result is the last checkpoint (step 760,000). The numbers on the two cards below are measured on early checkpoints with the board protocol:
training in progress; not a result.
CANONICAL NUMBERS (Glint-1.3
benchmark.py protocol):
64M v1 eff 75.81 (BLiMP 75.83 / ARC-E 47.94 / byte_ppl 2.372) ·
128M v1 eff 75.84 (77.26 / 51.94 / 2.272) ·
32M eff 75.41 (73.48 / 44.44 / 2.5386) ·
16M eff 74.33 (70.08 / 40.91 / 2.6746). The board still lists earlier values for our 64M (77.51); those came from a wrong bytes-per-token constant and a non-canonical BLiMP run.
GOAL: #1 EFF — measured with the same protocol, the strongest model we could evaluate (GPT-X2-125M) scores eff 78.26, about 2.4 above our v1 models (75.81 / 75.84); the gap is mostly ARC-Easy, with BLiMP second.
· board:
Glint Tiny-ML