Datasets:
Paired Spatial Information and Reader Diagnostics
Part of the Spatial Representation and Reasoning series. Companion dataset: SueMintony/spatial-representation-qa.
384 questions in one dataset: 96 paired primary, 240 invariant controls, and 48 epistemic/unknown controls. The remaining 288 controls are not a separate benchmark.
Organization update — 2026-11-10
Renamed from SueMintony/spatial-information-v4-pilot-20261110. Original question files, labels, images, frozen reader inputs, historical results and Git history are retained. Original split membership and question IDs are unchanged. The old dataset card is preserved in README.before-organization-c7bb1ba.md. This update adds offline audit views; it does not rerun models or certify fresh simulator/perception results.
| View | Rows | Use |
|---|---|---|
questions |
384 | Original question data; unchanged |
information_checks |
384 | Per-question role, GT basis, condition eligibility and modification reasons |
input_audits |
2688 | Actual operand/evidence checks and input-specific qualification |
reasoning_exam |
288 | Bounded offline eligibility manifest; references existing frozen inputs, not new questions or results |
Nested audit values are JSON strings to preserve their original types without inconsistent Arrow schemas. Audit views contain evaluator references: do not append whole audit rows to solver inputs. Read the referenced frozen reader_inputs for an actual model request.
信息检查与推理考试分别组织:先核对必要事实、时间、坐标系和原生精度,再判断是否有充分信息完成计算。控制题、直接检索、缺信息和待验证分别标注;答对但必要事实不一致不算表征事实正确。题号、配对与历史分数均保留,不按模型输赢筛题。
eligible_bounded_offline is limited to archived simulator JSON (Full100) or accepted visible-oracle annotations (paired diagnostics). Conservative readout unknown is not automatically proof of impossibility. Existing splits were engineering-inspected, not a new blind test. For comparative scores, freeze a shared sufficient manifest and retain the original coverage denominator.
Question versions remain explicit. For paired diagnostics, historical requests/results use v4-frozen-20261110; v4-plain-20261110-r1 English/Chinese wording appears only in audit review columns and has no new model scores. No original question wording is overwritten by this organization update.
For new runs, the response contract should explicitly allow direction labels such as east/west. This prospective clarification does not change historical requests or regrade their scores.
Machine-readable organization/provenance
Historical release documentation
The original release notes and scores follow. Their original scope, versions and limitations continue to apply.
Spatial Information v4 Paired Mechanism Pilot
24 paired worlds, 48 actual MuJoCo observation histories, 1,152 RGB-D frames, and 384 questions. Four newly authored single-floor layouts, three intervention programs, and two starts per configuration. This release organizes the frozen November 10, 2026 pilot without changing pixels, questions, labels or scores.
Read the QC report and browse images and text. Data packaging can pass while research scope remains limited. This is not a blind held-out benchmark or a full ToS/SATURN method ranking.
Configurations
| Config | Development | Validation | Contents |
|---|---|---|---|
| questions | 192 | 192 | Question + public context + visible-oracle annotation history; ordered embedded RGB evidence images. |
| observations | 576 | 576 | One RGB image, accepted annotation text, time, calibration and original RGB-D/mask file reference. |
| trajectories | 24 | 24 | 48 full histories, frame order and original public-text paths. |
| representations | 120 | 120 | Five question-independent retained views per trajectory: M/Q/R/T/H. |
| reader_inputs | 336 | 336 | Exact 672 original model requests; all are text-only. |
| labels | 192 | 192 | Evaluator-only gold answers, joined by query_id; JSON boolean/string/null types. |
| reader_results | 336 | 336 | 672 unchanged model outcomes and returned content; unknown/malformed remain in the denominator. |
| native_results | 6720 | 6720 | 13,440 deterministic interval/graph query outcomes over 35 conditions. |
| pairs | 12 | 12 | 24 world-pair mappings and query-pair identities; bookkeeping only. |
Load
from datasets import load_dataset
bundle = "/path/to/spatial_v4_dataset_20261110_001"
qa = load_dataset(bundle, name="questions", split="validation")
sample = qa[0]
model_text = sample["input_text"]
evidence_images = sample["images"] # Ordered PIL images.
# Pass these fields only, not the entire metadata row.
import json
requests = load_dataset(bundle, name="reader_inputs", split="validation")
formal_messages = json.loads(requests[0]["messages_json"])
# Exact original messages; the formal model received no image.
After uploading, replace bundle with the dataset repository ID. Every configuration also has JSONL mirrors. In JSONL, image values are relative paths; Parquet embeds original JPEG bytes and Hugging Face Image feature metadata.
Image and text boundary
The questions configuration is an audit-friendly full-observation track. Its text includes question semantics, public room/pose conventions and visible-only oracle object/doorway annotations. RGB images are ordered evidence companions. This multimodal packaging is not a new model evaluation. All 672 formal requests in reader_inputs used structured text only. An RGB-only replacement is not validated: opaque entity IDs require explicit identity grounding.
Original public histories are under assets/histories/<world_id>/public/. Each NPZ has depth_m in metres, entity_mask, and geom_ids, all 256 x 256. Image/camera calibration, own pose, times, visible identity anchors, and doorway observations are in the accompanying text. No hidden object anchor, hidden active graph, future event, evaluator label or provider answer belongs in model input.
Bookkeeping fields (query_id, pair/category/role/intervention/layout labels) are for indexing only. Use input_text and the intended image field, or the exact messages_json, rather than serializing a whole dataset row. Labels and scored responses are separate configurations. Gold is released for audit; this is not a protected future test set.
Question categories
| Category | All | Primary | Invariant | Unknown control | Meaning |
|---|---|---|---|---|---|
| precise_distance | 48 | 16 | 32 | 0 | Current horizontal anchor distance <= 0.925 m; stresses sub-metre precision. |
| current_connectivity | 48 | 16 | 32 | 0 | Current observed aperture connectivity under a room-waypoint route budget. |
| last_observed_region | 48 | 16 | 32 | 0 | Most recently observed entity room at or before the specified cutoff. |
| composed_visit_budget | 48 | 48 | 0 | 0 | Historical start room + current aperture route + current in-room metric approach. |
| current_room | 48 | 0 | 48 | 0 | Registered room containing the current observed target anchor. |
| coarse_distance | 48 | 0 | 48 | 0 | Current horizontal anchor distance <= 2.1 m; all labels are true in this pilot. |
| unobserved_entity | 48 | 0 | 0 | 48 | Return unknown for an entity absent from accepted observations. |
| world_direction | 48 | 0 | 48 | 0 | East/west ordering of target and reference; all labels are east in this pilot. |
Eight question categories are distinct from three interventions, six typed operators, five representations, three evaluation roles and four layout families. See metadata/categories.json, metadata/category_distribution.csv and metadata/input_contract.json.
Coverage limits: coarse-distance labels are all true; direction labels are all east; unobserved labels are intentionally all unknown. These are consistency/abstention controls, not balanced standalone capability evaluations. The 96 primary instances include 16 each of precision/connectivity/history atomic witnesses and 48 composed witnesses. One atomic and one composed question per history produce 48 query-pair witnesses over 24 world pairs.
Representations and conditions
- M: precise current horizontal object centres from a pinned official SATURN Scene component.
- Q: recovered project ToS global integer-map component at one metre per unit.
- R: current qualitative project graph (rooms, direction, coarse distance).
- T: observed current aperture connectivity, including observed blocked apertures.
- H: timestamped observed object-room transitions.
M/Q are component adapters, not full official systems. SATURN source pin: 409657d56c3b56055f4660dfb74332e2fe42c469. R/T/H are project views. The deterministic track has all 32 coalitions plus naive union, compact MTH and full annotation history. The model track has empty/Q/M/MQ/naive_concat/structured_fusion_MTH/history_only, each on 96 primary questions.
Q/R reconstructability is a designed structural redundancy control. Necessity certificates are conditional on this interface/query family. The deterministic reader is conservative, and its information-availability results must be separated from bounded-model utilization and lucky answers on ambiguous packets.
QC and research readiness
The source run passed 17,124 programmatic checks, 384/384 independent label calculations, 10 contract tests and eight exact capture reconstructions. Original preservation verified 55,528 prior files unchanged. Source receipts are copied under qc/original/; package-specific checks and limitations are in qc/PACKAGE_QC.json and QC_REPORT.html.
There are four architecture signatures, eight matched acquisition paths and four RGB hashes shared across development/validation. Development uses linear/ring; validation uses branch/ladder. All pair siblings stay together. Both groups were inspected for engineering. No blind test, visually disjoint split or broad building generalization is claimed.
Exact visible-only oracle identities/centres and own pose isolate encoding. The camera is a checked finite-radius kinematic proxy; changes are exogenous. There is no learned perception, SLAM, autonomous exploration, robot locomotion or manipulation result. Route cost is room-waypoint travel plus an in-room Euclidean approach, not a continuous shortest path.
The single DeepSeek reader made one attempt per input; 62/672 outputs failed the format contract. Provider routing was not fixed. No human QC sign-off is claimed. Full-method superiority and LLM-equivalence gates remain unvalidated.
Before scaling
Keep this pilot frozen. Create a new version with counterbalanced direction axes and distance thresholds, more object identities, more independent layouts and paths, and newly blind test groups. Stratify by category, intervention, pair and answer label; add human sample review. Register a repeated, provider-controlled common-reader experiment and preserve format failures. A learned-perception/RGB track additionally needs explicit entity grounding and separate perception QC.
Provenance and license
Original input pixels/text are copied byte-for-byte; provenance/COPIED_FILES.json records the mapping. Original source/data/model seal identities are retained under provenance/original_seals/. The independent package QC checks copied hashes, embedded images, typed labels and actual requests. No inference was rerun for this packaging step. No personal credential configuration or credential file is included.
New generated data and release helpers follow the project MIT license (see LICENSE). No third-party encoder implementation or model weights are redistributed. Repository: https://github.com/HLR/Dynamic-Embodied-Planning. This is a research pilot and must be cited/described with the scope restrictions above.
Hugging Face format references: image datasets, dataset cards.
- Downloads last month
- 802