SHAPE-Morph/benchmark-logs

SHAPE — Benchmark Run Logs

Full run logs (traces, per-image masks/metrics, deliverables, run summaries) for the SHAPE agent on two NeurIPS22-CellSeg benchmark tasks. Each archive is one full-1000-image run for one LLM backend, packed with tar + zstd -19 (.tar.zst).

Layout

logs/
  task1_seg/     # Task 1 — cell instance segmentation (NeurIPS22-CellSeg, n=1000)
  task2_crop/    # Task 2 — single-cell cropper (crops on GT masks, n=1000)
  task3_SSL/     # Task 3 — self-supervised representation learning (RxRx1 HEPG2 / RPE, RxRx3-core HUVEC)

task1_seg/ — instance segmentation

archive	LLM backend	macro AP50
2026-09-17_112053_concurrent-luna-final.tar.zst	gpt-5.6-luna	0.829
2026-09-17_210102_concurrent-solar-final.tar.zst	gpt-5.6-sol	0.830
2026-09-19_113551_concurrent-sonnet-5.tar.zst	claude-sonnet-5	0.822

Each archive contains, per image (images/<id>/): trace.json (agent+VLRR trace), metrics.json (instance-AP), deliverable_seg.png (raw | prediction | pred=cyan/GT=magenta), and shape_cache/ with every VLRR attempt mask (cell_mask_<id>_a{1..N}.npy) and vlrr_trace.json. Run-level: cell_image_metrics.csv, run_summary.json, run_params.json.

task2_crop/ — single-cell cropper

archive	LLM backend
2026-09-19_114615_concurrent-gt-luna-final.tar.zst	gpt-5.6-luna
2026-09-19_114623_concurrent-gt-solar-final.tar.zst	gpt-5.6-sol
2026-09-19_125305_concurrent-gt-sonnet5-final.tar.zst	claude-sonnet-5

Two arms per image: maskcrop/ (naive tight-bbox reference) and shape/ (agentic VLRR crop with in-loop triage VLM). Per image: shape/tool_cache/ with cell_crops_triage_trace_*.json (one per VLRR round) and crop metadata. Run-level CSV carries contamination_maskcrop, contamination_shape, contamination_reduction, retention_*, shape_vlrr_status.

task3_SSL/ — self-supervised representation learning

SHAPE-ES vs ReAct (at SHAPE-ES's control frequency), each with gpt-5.6-luna and claude-sonnet-5-5, against no-LLM baselines: fixed-recipe and random searches over the same levers, the default DINOv3 fine-tune, and frozen pretrained models (DINOv3, DINO4Cells, DCM, OpenPhenom, Phenom, CellProfiler). Training used a 25 % subset of each plate's training wells (seed 0). Scores on held-out wells: RxRx1 well mAP (HEPG2, RPE), RxRx3-core EFAAR recall (HUVEC).

archive	arm	HEPG2	RPE	HUVEC
shape_es_<task>_s0_luna	SHAPE-ES · gpt-5.6-luna	0.0184	0.0206	0.295
shape_es_<task>_s0_sonnet	SHAPE-ES · claude-sonnet-5-5	0.0215	0.0335	0.295
react_<task>_s0_luna	ReAct · gpt-5.6-luna	0.0116	0.0040	0.220
react_<task>_s0_sonnet	ReAct · claude-sonnet-5-5	0.0021	0.0151	0.260
policy_fixed_<task>_s0	fixed recipe, no LLM	0.0137	0.0331	0.300
policy_random_<task>_s0	random search, no LLM	0.0147	0.0380	0.287
default_<task>_s0	default fine-tune	0.0086	0.0198	0.262
frozen_models	frozen DINOv3	0.0145	0.0237	0.293
frozen_models	frozen DINO4Cells-S / -B	0.0150 / 0.0126	0.0295 / 0.0240	0.265 / 0.280
frozen_models	frozen DCM DINOv2-B	0.0072	0.0168	0.268
frozen_models	frozen OpenPhenom	0.0058	0.0090	0.267
frozen_models	Phenom-1 / Phenom-2 / CellProfiler (released tables)	—	—	0.318 / 0.358 / 0.286

Each run archive holds the agent trace (agent_log.jsonl), every LLM request and reply (reasoner.jsonl), every tool call with its training record (tool_calls.jsonl), run settings, the delivered model's record, per-call loss curves, the console log and the score; no checkpoints. Frozen models train nothing, so frozen_models holds their scores, scorer outputs and embedding code instead. Details: logs/task3_SSL/README.md. Code to rerun: SHAPE-Revision repro-sweep.

Unpack

tar --use-compress-program=unzstd -xf <archive>.tar.zst
# or: zstd -d <archive>.tar.zst && tar -xf <archive>.tar