scheduling pr follow-ups

[timurcarstensen]

TODO

• use yaml instead of json
• ~~put .sif images on fast ssd-based project scratch instead of hdd ~~
• stop duplicating .sif files in the workflow
• remove unnecessary HF env vars
• we have /shared_evals, /oellm_shared_evals, /evals, can we unify? Perhaps you could have eval_base_dir be the parent of what is currently specified and always add "oellm_shared_evals" which is defined in code.
• add other env vars for training workflow
• move container images and workflow to use OpenEuroLLM/evaluation_singularity_images
• use fine-grained HF_TOKEN in GH secrets for workflow that only has write access to OpenEuroLLM/evaluation_singularity_images

[geoalgo]

• use "|" to split tasks to allow passing passing model args such as EleutherAI/pythia-160m,revision=step100000
• rename PYTHON_PATH in the yaml to something like PYTHON_ENV as PYTHON_PATH is an environment variable with a special meaning
• do not use shell=True which is insecure

Also I was wondering what would be a good way to test, for instance to launch pythia evals on 3 clusters?