Natural Language Understanding on NLU Benchmarks (CSQA, HellaSwag, PIQA, SIQA, ARC Easy) 5-shot
66.2CSQA AccuracyHF/SMOLLM2-1.7B*
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| HF/SMOLLM2-1.7B*backbone_size=1.7B, pre-training_corpus=2T-11T tokens, evaluation_protocol=5-shot2026.07 | 66.2 | 71.6 | 78.2 | 53.2 | 80.1 | 69.9 | |
| HF/SMOLLM2-360M*backbone_size=360M, pre-training_corpus=2T-11T tokens, evaluation_protocol=5-shot2026.07 | 57.8 | 55.9 | 72.6 | 48.2 | 72.7 | 61.4 | |
| CO-LMLM-360M-FWbackbone_size=360M, pre-training_corpus=Fewer than 100B tokens, evaluation_protocol=5-shot2026.07 | 48.3 | 46.9 | 68.9 | 45.2 | 66.1 | 55.1 | |
| HF/SMOLLM2-135M*backbone_size=135M, pre-training_corpus=2T-11T tokens, evaluation_protocol=5-shot2026.07 | 46.3 | 42.4 | 67.6 | 45.6 | 66.9 | 53.8 | |
| STANDARD-360M-FWbackbone_size=360M, pre-training_corpus=Fewer than 100B tokens, evaluation_protocol=5-shot2026.07 | 46 | 47.3 | 69.4 | 44 | 68.4 | 55 | |
| STANDARD-360Mbackbone_size=360M, pre-training_corpus=Wikipedia, evaluation_protocol=5-shot2026.07 | 32 | 31.6 | 59.5 | 41.4 | 44.6 | 41.8 | |
| REL-LMLM-360Mbackbone_size=360M, evaluation_protocol=5-shot2026.07 | 31.7 | 32.9 | 59.7 | 41.3 | 45 | 42.1 | |
| REL-LMLM-135Mbackbone_size=135M, evaluation_protocol=5-shot2026.07 | 29.9 | 29.6 | 57.3 | 40.6 | 41 | 39.7 | |
| CO-LMLM-360Mbackbone_size=360M, pre-training_corpus=Wikipedia, evaluation_protocol=5-shot2026.07 | 29.4 | 31.9 | 57.1 | 40.5 | 44.3 | 40.6 | |
| STANDARD-135Mbackbone_size=135M, pre-training_corpus=Wikipedia, evaluation_protocol=5-shot2026.07 | 27.1 | 28.7 | 55.9 | 40.6 | 39.8 | 38.4 | |
| CO-LMLM-135Mbackbone_size=135M, pre-training_corpus=Wikipedia, evaluation_protocol=5-shot2026.07 | 26.7 | 29.2 | 56 | 40.2 | 40.1 | 38.5 | |
| Random Chanceevaluation_protocol=5-shot2026.07 | 20 | 25 | 50 | 33.3 | 25 | 30.7 |