Retrain extractor using llama 2, minstrel, or some other LLM

[arjunpatel7]

Current LLM parser uses Bloom, which is now been surpassed by other LLM models. Would love to substitute this model with llama 2, minstrel, or some other instruction fine-tuned LLM.

[arjunpatel7]

The following needs to be done:

• Identify a suitable LLM, preferably at or under 7b
• Set up Weights and Biases account to track LLM finetuning
• Rewrite LLM training script to work on modal, or google colab
• Regenerate datasets for this model
• write some additional loss functions/checks that estimate model performance on parsing different ways, including percentage of dictionaries that are parseable, names that are correctly parsed, numbers, etc
• model versioning, to test improvements and such
• Automate the script running/logging process, so that these loss results, and metrics, and models are saved (probably with weights and biases)
• Rewrite any scripts necessary to adjust for new LLM in use, probably in the inference layer for modal and the app
• Redeploy model app to use new LLM

Yields:

• 1 or more fine-tuned LLMs that we can hot-swap with the current Bloom model.

[arjunpatel7]

It may be useful to wrap said script into some sort of Modal command line utility, so we can point it at a dataset and retrain, rename, and track models at will. So, when we get around to doing the damage calculations in addition to this stuff, it'll be super easy to train the models without having to do too much