| model | avg | arc | truthfulqa | hellaswag | mmlu |
|---|
| crumb/d1536-250MT-full | 30.61 | 23.29 | 47.1 | 26.60 | 25.45787898435144 |
trained for 250MT/16GT, compared to GPT-2 suite: between gpt2-124m and gpt2-345m (the params are between as well). 254,883,841 non-embedding parameters, this is a roughly 1Tok/Param initial saturation of weights to be further pretrained with ReLora.
