Updates for regenerating text

[MinaAlmasi]

I would suggest the following steps for when we want to rerun the text generation with new models:

Do before re-generating

• Update vLLM and other necessary packages, so we can also update the Python version.

  Currently running everything with coder Python 1.87.2 on UCloud which has Python 3.10. There have been 9 updates since then to the UCloud App.

• Look into whether vLLM have added a "min_token" parameter.

  Currently, I compute the length of strings and re-generate in a for loop for n = 20 times to avoid getting generations below the desired amount of tokens for each task. There is no need for this hacky solution if there is a built-in solution now.

• Consider using entire model-names instead of the current short-hands

  E.g. use stabilityai/StableBeluga-7B or StableBeluga-7B instead of beluga7b

• Remove model names as prefixes for completions column

  Back when I started the project, I somehow thought it was a good idea to add the model name as prefix to the "beluga7b_completions" which I ultimately remove in the folder make_dataset to standardise formats across models. It should just be called completions

After re-generating

• Remove HF pipeline

  At the time of coding, I also created the possibility of using the HF interface also. I think for simplicity we should remove this. There is no need for the scope of the project (esp. if we want to split the repos at some point).

• Run embeddings with smaller model

  I used nvidia/NV-Embed-v2 because it scored the highest on MTEB, but it is a heavy model - is it overkill for a baseline? I changed from FP32 to FP16 precision to make it less memory hungry, and could run it with a batch-size of 16 on the new nvidia L40 gpus