831 B
831 B
SmolLM Target Benchmark
| Task | Metric | ours-stage2 | smollm-135m | smollm2-135m |
|---|---|---|---|---|
| ARC-Challenge | acc_norm,none | 0.2807 | 0.2875 | 0.2969 |
| ARC-Easy | acc_norm,none | 0.5059 | 0.5610 | 0.5854 |
| HellaSwag | acc_norm,none | 0.3242 | 0.4265 | 0.4301 |
| LAMBADA OpenAI | acc,none | 0.3757 | 0.4289 | |
| PIQA | acc_norm,none | 0.6137 | 0.6823 | 0.6839 |
| WinoGrande | acc,none | 0.5272 | 0.5249 |
Means
ours-stage2: 0.4311smollm-135m: 0.4767smollm2-135m: 0.4917
Candidate Gaps
ours-stage2vssmollm-135mmean gap: -0.0456- HellaSwag: -0.1023
- PIQA: -0.0686
- ARC-Easy: -0.0551
- ARC-Challenge: -0.0068
ours-stage2vssmollm2-135mmean gap: -0.0605- HellaSwag: -0.1059
- ARC-Easy: -0.0795
- PIQA: -0.0702
- ARC-Challenge: -0.0162