vineet-datasets/probe-benchmark-600

approved-questions/ — PROBE-Bench V2 benchmark curation & handoff

Latest benchmark finalized: September 10, 2026 (Pacific time). See benchmark_600/README.md for the authoritative current release, human-review changes, and HF rollback instructions. Find/beneath edits are baked; 68 compare questions use their accepted inverse. Current compare polarity is 77 largest / 73 smallest; beneath has 39 NOTA (26%); find is 75 yes / 75 no. Count's first-round edits remain. The curation notes below describe the original selection and contain historical distributions/runtime commands; bank/ and leftover_pool/ remain archives.

Read this fully. This folder is the single source of truth for the finalized PROBE-Bench V2 benchmark and everything a follow-up agent needs to (a) deploy the review website, (b) swap questions, (c) regenerate/extend the benchmark with run_v2.py, or (d) re-derive the selection from scratch.

⚠️ Git-ignored (approved-questions/ is in .gitignore). ~14 GB total, far above the 5 MB commit limit. This is a local working artifact on vineetb-desktop (repo root /home/vineetb/Desktop/projects/Agentic_MQA). Do NOT try to commit it.

---

TL;DR — what's here

Path	What	Size	Use
benchmark_600/	THE DELIVERABLE — frozen 600Q benchmark (150/type)	~5 GB	--from_snapshot eval; review website
leftover_pool/	240 accepted questions NOT in the 600	~2.7 GB	swap-in reserve
bank/	ALL 783 reviewer-accepted questions (superset)	~6.5 GB	full archive / re-selection
POOL_STATS.txt	distributions for all three banks	—	at-a-glance stats
scripts/	the exact scripts that built everything	—	reproduce / extend

All three banks are --from_snapshot-replayable rollout-shaped dirs (renumbered ep_NNN_<type>/ + bank_manifest.json + _ref_views/ + _texture_store/).

---

1. The frozen benchmark — benchmark_600/

600 questions, 150 per type. Curated to maximize object diversity (the explicit design goal), not to balance ground-truth.

# Evaluate a model on the frozen benchmark (scores directly — GT is in each
# snapshot's annotation, so no question re-generation happens):
python run_v2.py --mode agentic --model_name <MODEL> \
    --from_snapshot approved-questions/benchmark_600 --n_workers 4
# singlepass baseline:
python run_v2.py --mode singlepass --model_name <MODEL> \
    --from_snapshot approved-questions/benchmark_600 --n_workers 4

Composition

Type	N	Detail
count	150	116 yes (GT≥1) + 34 no (GT=0). GT dist {0:34,1:12,2:24,3:24,4:34,5:22}. ≤3 questions per counted-object.
find	150	97 yes + 53 no. ≤3 questions per searched-object.
beneath	150	all-unique (cover, buried-target) duplets. GT = revealed object name; 26 "none of the above" (~17%).
compare	150	balanced 75 largest / 75 smallest; all-unique 3-object triplets. Letters A:53/B:57/C:40.

Diversity achieved

• 319 unique object folders = 100% of the whole accepted pool's variety.
• count/find: 71 distinct asked-objects each, capped at ≤3 reuse. This is near the theoretical floor: the EVAL asset pool only has 82 objects with role='target', and count/find can only ask about target-role objects.
• beneath/compare: every signature (duplet / triplet) is unique → 150 distinct each.
• Scene distractors were greedily chosen (max new-object coverage) sharing ONE global object set across all 600 questions.

Selection strategy (agreed with Vineet)

• count/find — "Option A": per-asked-object cap of 3 over yes+no combined. YES from the reviewed pool; NO/GT=0 fill pulled from source rollouts (count_v2, find_v2, qbank). NO/GT=0 questions need no human review because the asked-object is absent by construction → GT=0 / "no" is algorithmically guaranteed. This let us drop the yes-cap to 3 and still hit 150.
• compare — rebalanced away from the pool's natural 59%-smallest skew to an even 75/75. Unique triplets ⇒ zero diversity cost.
• beneath — greedy max object-coverage over unique duplets; NOTA left natural.

Layout (per episode)

benchmark_600/
├── bank_manifest.json     authoritative 600-episode list (SEE §4)
├── _ref_views/            object reference orbits, incl. badged compare markers
│                          (<source>__<folder>[__badge_<hash>]/N.jpg)
├── _texture_store/        table-texture blobs referenced by snapshots (by hash)
└── ep_000..ep_599_<type>/
    ├── scene_snapshot/scene_snapshot.json   REPLAYABLE poses + annotation
    │                                        (annotation.question_descriptor = GT)
    ├── review.json                          question + GT + reviewer verdict
    ├── verification.png                     faded-clutter GT-reveal image
    ├── after_walls_topdown.png              CLEAN review frame (no badges/markers)
    ├── *_gt_scene.png                       GT bbox overlay (debug)
    ├── *_seg_overlay.png                    segmentation overlay
    └── *_occlusion.json / gt_visibility.json / selected_objects.json

---

2. The reserve — leftover_pool/

240 accepted questions not selected (beneath 50, compare 69, count 58, find 63). Zero overlap with benchmark_600 (disjoint source_episodes). Use these to swap in replacements if collaborators reject benchmark questions during the screening pass.

• Same rollout-shaped layout + manifest.
• Its count/find here have HIGHER per-object reuse than the benchmark (they're the leftovers), so prefer picking leftover episodes whose asked-object is NOT already at the ≤3 cap in the benchmark. scripts/diversity_analysis.py + scripts/plan_options.py show how to recompute reuse.
• compare leftovers are extremum-skewed (55 smallest / 14 largest) — if you swap a compare question, keep the 75/75 balance in mind.
• NO/GT=0 count/find fill can ALSO be regenerated cheaply from the source rollouts (they hold ~330 count-GT0 + ~566 find-no episodes, mostly unreviewed).

---

3. The full archive — bank/

All 783 reviewer-accepted questions (beneath 200, compare 219, count 181, find 183) = benchmark_600 ∪ leftover_pool MINUS the NO-fill episodes that came from unreviewed sources. Built by scripts/finalize_bank.py --per-type 0.

• bank/bank_manifest.json additionally flags 6 "review-tab-hidden" episodes (review_tab_hidden=true): 3 count GT=0 + 3 find-"no" that the live review TABS hid via display filters (:+nogt0 / :+yesonly) but were still reviewer-accepted. accepted_total=783, accepted_served_pool=777.
• Compare inverse verdicts: compare_inverse_includable=150 (150 of 219 accepted compare questions have an includable largest↔smallest inverse; each such inverse flips polarity).

To re-derive a DIFFERENT selection (e.g. different caps, or 700Q), edit knobs at the top of scripts/finalize_600.py (N, CAP, per-type logic) and re-run.

---

4. bank_manifest.json schema (all three banks)

{
  "schema_version": 1,
  "description": "...",
  "seed": 1000000,
  "n_episodes": 600,
  "by_type": {"beneath":150,"compare":150,"count":150,"find":150},
  "episodes": [
    {
      "ep_dir": "ep_000_beneath",       // dir under the bank root
      "question_type": "beneath",        // count|find|beneath|compare
      "gt_answer": "white skincare box", // resolved GT (post reviewer edit)
      "polarity": "yes"|"no"|null,       // count/find only (yes=GT>=1, no=GT=0)
      "extremum": "largest"|"smallest"|null,  // compare only
      "source_rollout": "rollout_vlm_20260827_231647",  // provenance
      "source_episode": "ep_1633_beneath"              // original dir name
    }
  ]
}

discover_snapshot_episodes() (in core/scene_snapshot.py) honours a bank_manifest.json if present (pins order/subset); otherwise it globs ep_*/scene_snapshot/. Either way --from_snapshot works.

GT-leak note: review.json NULLS target_obj_id in its question block for safety. The authoritative target_obj_id lives in the snapshot annotation and in *_question.json. Read tid from there, never from review.json.

---

5. Review WEBSITE deployment

The collaborative review server is scripts/review_server.py, launched via start_review_server.sh (repo root). It currently serves TWO tabs:

benchmark_600  → approved-questions/benchmark_600   (the deliverable; 0 reviewed = fresh)
qbank          → rollouts/rollout_vlm_20260827_231647 (original 1726-bank; 949 reviewed)

The old count_v2, find_v2, mika_v1, mika_v2 tabs were retired once the benchmark was finalized (their source rollouts + review decisions remain on disk). To bring one back, re-add its --bank line.

Hard harness constraint — server persistence

Any process backgrounded inside an agent tool call is KILLED when the call returns. The USER must launch the server in their own persistent terminal:

tmux new -s review
./start_review_server.sh            # binds 0.0.0.0:8090
# Ctrl-b d to detach; tmux attach -t review to return

Browser: http://10.20.13.172:8090/ (LAN) or localhost:8090. After ANY change to tabs/filters/episodes, restart the server + hard-refresh (Ctrl-Shift-R) — the episode list is discovered ONCE at startup and cached in the BANKS dict for the server's lifetime.

Firewall: 8090/tcp may still need sudo ufw allow 8090/tcp (needs user sudo).

Bank CLI spec + filters

id:Display Name:/path/to/rollout[:+nogt0][:+yesonly]

• :+nogt0 — hide GT=0 episodes (non-destructive display filter).
• :+yesonly — for yes/no banks, show only gt=='yes'.
• <rollout>/EXCLUDE_EPISODES.txt — always-hidden hold-out list (one ep name per line, # comments OK). The benchmark_600 tab uses NONE of these (it's the final curated set — everything should be shown).

What the review UI supports (all question types)

Accept / Reject / Ambiguous, plus edits: GT edit (mandatory reason), options edit, question-text edit, and (compare) inverse-question annotation. Decisions append to <bank>/reviews/<reviewer>.jsonl (last-write-wins per episode; claims.jsonl = soft-locks, skip it when aggregating). Reference-object guides render from _ref_views/ (that's why each bank ships its own _ref_views/).

If a reviewer edits GT on benchmark_600

Edits land in benchmark_600/ep_*/review.json AND (for a frozen bank) should be propagated into the snapshot annotation before scoring. scripts/finalize_bank.py does this automatically when building FROM a reviewed rollout; if editing the already-frozen bank in place, re-run the GT-edit propagation (see _apply_gt_edit / _apply_text_edits in scripts/finalize_bank.py).

---

6. Provenance — where questions came from

Bank tag	Source rollout	Role
qbank	rollouts/rollout_vlm_20260827_231647	original 1726-candidate bank (all 4 types)
count_v2	rollouts/rollout_vlm_20260901_220948	occlusion-forced count, seed 3000000
find_v2	rollouts/rollout_vlm_20260902_044842	occlusion-forced find + yes top-up (merged)

These source rollouts also under rollouts/ (git-ignored, safe to keep). The NO/GT=0 count/find fill in benchmark_600 was drawn DIRECTLY from these source rollouts (mostly unreviewed episodes — legitimate because GT is algorithmic). Nothing in the sources was modified by any finalize step.

---

7. Regenerating / extending questions with run_v2.py

Snapshot-bank GENERATION (make NEW candidate scenes; no VLM scoring):

# generate NEW scenes only (snapshots), e.g. more find-"no" for a bigger pool:
python run_v2.py --mode agentic --model_name gcp/google/gemini-3-flash-preview \
    --snapshots_only 1 --question_type find --find_polarity no \
    --n_episodes 300 --seed 7000000 --n_workers 4

Key run_v2.py behaviours (see run_v2.py::_pin_v2_defaults):

• --from_snapshot <bank> → replays that bank, sizes n_episodes to the bank, scores against each snapshot's annotation. Mutually exclusive with --scene_bank.
• --snapshots_only 1 → generation mode; honours your --seed/--n_episodes /--question_type. --question_type {count,find,beneath,compare} forces all eps to that type; random → even split.
• --find_polarity {mixed,yes,no} (default mixed = even→yes/odd→no). no forces every find ep to "no" (absent asked-object). ⚠ count has no GT=0 forcing flag yet — count GT is rng.randint(0,5) in questions.py::_prepare_counting_scene; add one there + wire through core/worker.py + BOTH parsers if you need forced count-GT0 generation.
• DUAL PARSER PATTERN: any NEW CLI flag must be added in core/cli.py (worker parser) AND orchestrator/parallel.py (parse_args) AND forwarded to workers in parallel.py's cmd-build block, or --n_workers > 0 errors with "unrecognized arguments".
• Seeds already used (pick a provably-disjoint seed for fresh gen): eval frozen 1000000; qbank 1000000 (spanned up to ~2.8M); count_v2/find_v2 3000000; find-yes top-up 5000000. Use ≥ 7,000,000 for anything new.
• Worker recycling (--max_eps_per_subprocess, default 8) is the ONLY fix for the PyBullet C++ mesh leak (~700 MB/ep). Single-process (--n_workers 0) has no orchestrator to relaunch — set --max_eps_per_subprocess 0 there.
• Occlusion bar shared across V2 tasks: V2_OCCLUDED_MAX_VISIBILITY = 0.30 (core/utils.py).

After generating new candidates → build review artifacts → review → finalize

# 1. verification images (PyBullet software, ~7-8s/ep):
python scripts/build_verification_images.py rollouts/rollout_vlm_<TS> --only find
# 2. reference-view orbits (8-view; --all-scene-objects for full guides):
python scripts/build_reference_views.py rollouts/rollout_vlm_<TS> --only find --all-scene-objects
# 3. add a review tab (start_review_server.sh) → user reviews in tmux
# 4. freeze the reviewed rollout into a scored bank:
python scripts/finalize_bank.py rollouts/rollout_vlm_<TS> --per-type 150 \
    --out approved-questions/<new_bank>

---

8. scripts/ — the exact tooling that built this folder

Run all from the repo root (they hard-code repo-relative paths). Idempotent.

Script	Purpose
finalize_600.py	Builds benchmark_600/ + leftover_pool/ from bank/ + source-rollout NO-fill. Knobs at top: N=150, CAP=3, per-type logic (Option A for count/find; 75/75 compare; unique beneath). Re-run to change the selection (idempotent — rewrites both dirs). --dry-run prints the plan without writing. Copies _ref_views for BOTH scene-object folders AND the review.json reference block (compare marker refs carry a __badge_<hash> suffix absent from scene folders — this second copy pass is built in; don't drop it if you refactor, or the website 404s the compare referred-object guide).
finalize_bank.py (in repo scripts/, not here)	Builds bank/ (all-accepted) from reviewed rollouts; applies GT edits; copies texture store; writes manifest. --per-type N caps per type, 0=keep all.
collate_benchmark.py	The original accepted-count + inverse-pair + distribution report across the 3 review banks (respects :+nogt0/:+yesonly/EXCLUDE filters → gives the 777 served headline).
diversity_analysis.py	Per-type signature-reuse histogram + unique-object counts on bank/.
plan_options.py	Compares NO-fill options (cap 2/3, disjoint vs Option A) → the table that decided the strategy.
augment_manifest.py	Adds the review_tab_hidden flags + summary fields to bank/bank_manifest.json.
gen_pool_stats.py	Regenerates POOL_STATS.txt for all three banks.

To rebuild everything from the reviewed source rollouts

# (1) all-accepted archive:
python scripts/finalize_bank.py \
    rollouts/rollout_vlm_20260827_231647 \
    rollouts/rollout_vlm_20260901_220948 \
    rollouts/rollout_vlm_20260902_044842 \
    --min-accepts 1 --per-type 0 --out approved-questions/bank
python approved-questions/scripts/augment_manifest.py
# (2) frozen 600 + leftovers:
python approved-questions/scripts/finalize_600.py
# (3) stats:
python approved-questions/scripts/gen_pool_stats.py

---

9. Key facts a follow-up agent MUST know

• EVAL asset pool has only 82 role='target' objects → hard ceiling on distinct count/find asked-objects. 71 are already used across the accepted pool. To raise count/find diversity you must generate NO questions asking about the ~11 unused target objects (or promote more objects to target role in owg_robot/assets/object_annotations_eval.json).
• compare/beneath are already 100% signature-unique — never a diversity problem; only balance/count matter.
• NO/GT=0 count/find = review-free (algorithmic GT) → the cheapest way to grow those types.
• Object-folder key format: <source>_objects/<obj_name> (e.g. dtc_objects/dtc_figurine_b0cjf6mmg5_graycat). _ref_views dir key replaces /→__. Compare marker refs append __badge_<hash>.
• question_type in review.json for compare is marked_comparison (V1 name); canonicalize with {marked_comparison:compare, counting:count, existence:find, under_object:beneath}.
• Everything runs inside conda env agentic_mqa (Python 3.9.18); interpreter /home/vineetb/miniconda3/envs/agentic_mqa/bin/python.
• Test-command hygiene: filter stderr with grep -vE "pybullet build|Warning: get"; export MALLOC_ARENA_MAX=2.