experiments/README.md

Experiments archive

This directory contains dataset-specific runners, exploratory objectives, evaluation code, reports' source artifacts, external research repositories, the rollout viewer, and historical outputs. It is intentionally outside the final OURS library surface.

Experiment code imports the independent backbone from OURS/src/uniar. From the repository root, use:

export PYTHONPATH="$PWD/OURS/src:$PWD/experiments/src"

Keep OURS/src first so the upstream checkout cannot shadow the owned uniar package. Dataset-backed artifacts belong in experiments/output/<dataset>/<category>/<task>/YYYYMMDD-HHMMSS_<name>/. Use runs, sweeps, and smoke for new work; archive is reserved for failed, interrupted, duplicate, or superseded runs moved after termination.

The 2026-08-27 migration preserved every run basename and rewrote movable configuration/summary references to the new locations. Source-dataset paths inside provenance records remain unchanged.

cluster_jobs owns submission-time source snapshots and safe Slurm launch. New queued work uses this wrapper so edits made during queue wait cannot change the launched code. See docs/submission_snapshots.md for its contract and CLI.

qwen3_5/eval_visual_jigsaw.py is the resumable Qwen3.8-27B vLLM baseline for the complete 1,095-row Visual-Jigsaw test split. Submit it on one GH200 through qwen3_5/run_visual_jigsaw_vllm.sh; the runner uses the dedicated Qwen38-vLLM environment and the checkpoint chat template with its official reasoning_effort="medium" argument.