omg
#1
by Banaxi-Tech - opened
I had talked to him , There are a few models with 151 M & he said its fine its just only 1M param more .
OK
Same as of the Rose-Medium , mix of fineweb edu , cosmo & dclm
@GODELEV did you actually get 1.6259081413888932 loss or was 1.6259081413888932 for some kind of SFT
SFT ? Hell no No sft its just base model
ok thanks ill run v6.5 and add it
Banaxi-Tech changed discussion status to closed
@GODELEV theyre added, if you want detailss:
| Tasks | Version | Filter | n-shot | Metric | Value | Stderr |
|---|---|---|---|---|---|---|
| ValARC | 1.2 | none | 0 | acc | 0.6760 | 0.0105 |
| none | 0 | acc_norm | 0.5760 | 0.0111 | ||
| PhrasingARC | 1.2 | none | 0 | acc | 0.6700 | 0.0210 |
| none | 0 | acc_norm | 0.5740 | 0.0221 | ||
| ValHellaSwag | 1.2 | none | 0 | acc | 0.3945 | 0.0109 |
| none | 0 | acc_norm | 0.4100 | 0.0110 | ||
| PhrasingHellaSwag | 1.2 | none | 0 | acc | 0.3740 | 0.0217 |
| none | 0 | acc_norm | 0.4000 | 0.0219 | ||
| ValPIQA | 1.2 | none | 0 | acc | 0.6385 | 0.0107 |
| none | 0 | acc_norm | 0.6295 | 0.0108 | ||
| PhrasingPIQA | 1.2 | none | 0 | acc | 0.6640 | 0.0211 |
| none | 0 | acc_norm | 0.6400 | 0.0215 |
| v6.5-contam-detection | 1.2 | none | 0 | acc | 0.9840|