TrainLLMv3

framework for [eventually] achieving closed-book qa

#see 4-phase-training (old document) #atm only doing phase II. Phase I is meant to pre-train

#important files #resources/train.json #vars.py

#pre-training Generate_questions.py #writes outputted json to llm dataset format of contexts, questions, answers prep_raw_data.py

#other files of interest

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/user/env/lib/python3.11/site-packages/nvidia/cusparse/lib/ export WANDB_MODE=offline ./mezo-run.sh prep_data.py ./mezo-run.sh mezo-all-phases.py

#after training mezo-inference.py