Xnhyacinth/DT-Sched-Bench-core

DT-Sched-Bench Core Eval Data (data/)

This directory is the self-contained, HF-distributable eval bundle for DT-Sched-Bench v0.51.0. With data/ populated, the full 351-row core can be run without cloning the 8 upstream works/ git repos.

Layout

data/
  scenarios/      351 v0.51.0 core scenario YAMLs (flattened to
                  <domain>/<family>/<mode>/<level>/<seed_id>.yaml,
                  release-agnostic)
  release/        v0.51.0 manifest + registry/primary/core suite JSONs
  backends/       backend runtime data each released backend reads:
                    pglib_uc/  rts_gmlc/  pglib_opf/  opendss_ieee13/
                    jsplib/  pyvrp_instances/  or_gym/  sumo_ingolstadt/
                    nrel_microgrid/
  MANIFEST.json   SHA-256 of every tracked file + counts
  README.md       this file

How data/ relates to works/

• works/ is the developer/raw-clone location: full git repos at pinned commits, including .git history. Populated by scripts/setup_eval_env.sh.
• data/ is the curated mirror: only the files each backend actually reads, flattened, HF-uploadable. Built by scripts/build_data_dir.py.
• scripts/build_data_dir.py --link-backends symlinks works/<name> → data/backends/<name> so backend code that reads works/ resolves to the curated data/ copy (use this when you fetched data/ from HF and don't have works/).

Get this data

Option A — build locally (full upstream clones, recommended for development):

bash scripts/setup_eval_env.sh                 # installs deps + clones works/ + builds data/

Option B — download from HuggingFace (no git clones, faster for eval-only):

python scripts/download_from_hf.py             # fetches data/ from HF
bash scripts/setup_eval_env.sh --skip-data     # install pip deps only
python scripts/build_data_dir.py --link-backends   # symlink works/ -> data/backends/

Upload to HuggingFace

export HF_TOKEN=hf_...                         # from https://huggingface.co/settings/tokens
python scripts/upload_to_hf.py                 # uploads data/ -> Xnhyacinth/DT-Sched-Bench-core
python scripts/upload_to_hf.py --dry-run       # validate without uploading

Coverage (verified on 2026-07-21)

350/351 core rows runnable (99.7%). 13/14 backends pass a wait_only smoke episode (cigre, grid2op ×3 sub-envs, jsplib_job_shop, mock_sumo, opendss_fresh_feeders, opendss_ieee13, orgym_invmgmt, pandapower_acopf, pandapower_lv, pglib_uc_synthetic, pymgrid_economic_dispatch, pyvrp_cvrp, pyvrp_vrptw).

The one remaining gap:

• sumo live (1 row): gated on DT_TRAFFIC_BACKEND_REAL=1 + a reachable libsumo/traci transport.

License

Each backend dataset retains its upstream license (see docs/DATA_PROVENANCE.md for the per-dataset license table). The scenario YAMLs and release manifests are DT-Sched-Bench's own metadata, MIT-licensed with the benchmark.

SUMO live (1 row, opt-in)

The live SUMO backend needs eclipse-sumo (installed by requirements-backends.txt or pip install eclipse-sumo). Activate it with DT_TRAFFIC_BACKEND_REAL=1. The backend auto-detects the transport (traci/libsumo); no Docker or manual install needed.

pip install eclipse-sumo                            # one-time
DT_TRAFFIC_BACKEND_REAL=1 python run.py --scenario ... --agent wait_only --seed 42

Note: the 1 sumo core row (traffic_live_incident_response_deep_planning_basic) is a pilot row; the remaining 51 traffic rows use the deterministic mock_sumo backend which needs nothing beyond stdlib. The traci lifecycle has a known pre-existing issue with connection re-use across reset cycles (traci complains "Connection already active"); the transport detection fix in this PR makes sumo discoverable, but the 1-row live pilot may still need per-episode process cleanup.