Feedback on File Paths and Output Structure (Few-Shot Pipeline)

[SAM12302]

First, thank you for making this codebase and the few-shot SEAL pipeline available. It’s been genuinely useful to work through. While running the few-shot setup, I noticed a couple of small things that might be worth tightening for future users.

Notes

Path inconsistencies between scripts and README
In the few-shot pipeline, some scripts (for example, BC-self-edit.py) save outputs to fixed paths like ./RL_trained_model, while the README references paths such as RL_trained_model_iteration_1_8_epoch or RL_trained_model_iteration_1. This makes it a bit unclear which model directory is meant to be used in later steps.

Output files overwritten across runs
Files like final_results.json get overwritten by later stages of the pipeline, so repeating the few-shot experiments wipes out previous results.

Static folder structure for experiments
Model outputs, LoRA weights, and metadata all go to fixed directories, which makes it tricky to run multiple few-shot experiments or iterations without manual renaming.

Suggestion

It might help to make experiment outputs dynamic per run (or per iteration), and to unify the naming conventions between the README and the scripts. This would avoid overwrites and keep the few-shot workflow easier to follow.