1_raw_dataset/: upstream/raw dataset copies.
2-1_head_identification_preprocess/: scripts and converted JSONL files for head-identification datasets.
2-2_head_identification/: attention collection, head scoring, head sorting, visualization, and threshold sweeps.
3-1_model_training_preprocess/: training-data generation scripts and output directory.
3-2_model_training/: LoRA training scripts and output directory.
4_model_evaluation/: model evaluation and summary scripts.
sharedlibs/: shared Python helpers used by multiple stages.

Path Rules

Scripts under each stage use relative paths:

raw dataset inputs point to ../1_raw_dataset/...
head-identification inputs point to ../2-1_head_identification_preprocess/...
head-scoring outputs point to ../2-2_head_identification/head_scoring/...
generated training data points to ../3-1_model_training_preprocess/model_training/...
LoRA outputs point to ../3-2_model_training/lora/...
local model checkpoints point to ../../models/...

Each runnable stage has a lib symlink to ../sharedlibs/lib so existing from lib... imports continue to work.