README.md

PyINE scripts

Helper scripts we used during our own experiments. None of these are required to use the framework; they exist as reference launchers and sweep drivers. Open the individual files for full usage details (most have a --help block or a comment header).

A few rely on SLURM (flagged below); the rest run on any Linux box with bash + GPUs. For multi-node SLURM training in particular, see SLURM_GUIDE.md.

Launchers

• run_ddp.sh: torchrun wrapper that auto-detects GPU count, picks a free rendezvous port, and forwards Hydra overrides to a target app. Works single-node or under SLURM.
• launch_slurm.sh: SLURM sbatch script for multi-node DeepSpeed ZeRO-3 RL training with colocated vLLM rollouts.
• launch_debate_eval.sh: SLURM debate-eval launcher on a single 8-GPU node, with three modes (small vLLM interrogators, big vLLM interrogators with TP, and OpenAI-API interrogators).
• launch_wandb_agents.sh: SSHes into a list of GPU nodes and spawns 8 W&B sweep agents per node in tmux panes (one per GPU).
• my_command: one-line template invoked by launch_wandb_agents.sh to start a W&B agent inside each pane (substitutes REPO_ROOT, CUDA_DEVICE, SWEEP_ID).

Sweeps (sequential)

• run_checkpoint_sweep.sh: for each HF checkpoint in a list, spin up vLLM, run the hf_trainer predict pipeline, tear down.
• run_model_sweep.sh: same shape, but iterates over HuggingFace models instead of checkpoints.
• run_probe_sweep.sh: runs the probe trainer back-to-back across the weak / moderate / strong shortcut presets via run_ddp.sh.
• run_classifier_bias_sweep.sh: runs the LLM classifier trainer across the resampling-bias presets (weak / moderate / strong).
• run_prompted_llm_eval_openai_sweep.sh: sweeps the prompted-LLM guardrail eval across OpenAI models and reasoning-effort levels.
• run_prompted_llm_eval_vllm_sweep.sh: sweeps the prompted-LLM guardrail eval across a list of vLLM-served HuggingFace models.

Utilities

• check_ascii_staged.py: pre-commit hook that flags non-ASCII bytes on staged lines only (legacy non-ASCII content is tolerated).
• vllm_eval/: standalone vLLM serving + grading utilities for evaluation/judge workflows; see its own README.