Upload folder using huggingface_hub (#3)

Unloading models + current evals

reverted tokenizer fix, will be solved manually upon testing, current evals

Load tokenizer from parent model & app.py fixes

small python fix

gsm8k fix, queue, ref_model column

current eval

num_z = 50 for all

It was and not or 🙂

More aggressive OOM prevention, num_z reduced to 50 for gsm8k and mmlu

Trying to prevent OOM on mmlu and gsm8k by halving seq len

Avoiding re-loading already loaded models. Stated unload functionality as not-implemented.

Upload 16 files

First Test