yujl24/spatial-rsi-dynamic-mini

Spatial RSI Dynamic Mini

A deterministic multi-source research subset with full provenance. This is a reference manifest, not a bundle of all source images/videos or simulator worlds. Every question has pinned source metadata and media/environment configuration; do not run missing-media samples as text-only.

This is the single current public benchmark. Historical experiment snapshots remain internal for reproducibility. It retains the original seed records and their frozen splits, then adds category-stratified upstream evaluation samples using seed 20261012. New strata receive at most 2 Dev / 4 Eval records, with explicit smaller quotas for undersized strata. Selection never uses model scores or failures.

See RESEARCH.md for the expanded research catalogue, source coverage, inclusion decisions, scoring/input protocols and limitations. LICENSES.md records source-specific terms. Source releases are pinned in sources.lock.json.

Contents

Source	Dev	Eval	Total
ESI-Bench	58	116	174
EB-ALFRED	12	24	36
EB-Habitat	12	24	36
EB-Navigation	10	20	30
EmbodiedEval	10	20	30
EmbodiedNav-Bench	16	31	47
Total	118	235	353

Full metadata is in record_json and frozen dev.jsonl / eval.jsonl. Flat columns expose dataset, source repo/revision/sample/split, annotation/row SHA-256, original upstream dataset, category/subcategory, group, metric, runtime engine/status and observation count. Original questions, option mappings, native annotations, media locators, derived-category notes and sampling configuration are preserved. Target/hint/native state fields are evaluator-only.

Load

import json
from datasets import load_dataset
suite = load_dataset("yujl24/spatial-rsi-dynamic-mini")
record = json.loads(suite["dev"][0]["record_json"])
print(record["dataset"], record["source"], record["task"])

Dev is for prompt/spatial-skill optimization. Eval is a frozen local holdout, not a guarantee of unseen public data or absent pretraining contamination. Shared exact media/declared scene groups constrain splits, including historical seed assignments. Cross-source missing scene IDs and resized duplicates remain limitations; inspect group metadata before stronger generalization claims.

Media and runtime

Six source tracks / four benchmark families: ESI-Bench, EmbodiedBench (EB-ALFRED, EB-Navigation, EB-Habitat), EmbodiedEval and EmbodiedNav-Bench. They require online action–observation interaction; no reference demonstration trajectory is used as the model observation.

ESI preserves the existing integration. New source records contain native initialization metadata or pinned native episode locators, official selectors/indices and verifier descriptions. EB-ALFRED selected raw episodes and all selected ESI metadata are retained as evaluator-only annotations. Habitat numeric state references are resolved into object configurations/transforms. EmbodiedEval includes original task settings/predicates. EmbodiedNav includes initial drone pose and hidden target/reference route metadata.

The new simulators and policy bridges have not all been deployed on 175. runtime.deployment_status makes this explicit. Use each pinned upstream native simulator and its goal/predicate/navigation verifier before running these tracks. Simulator scenes/assets and dependencies are not bundled. Dataset construction/publication is independent of the running/frozen ESI Eval.

EmbodiedNav sampling uses derived vertical-displacement/route-length bins and disjoint 50 m horizontal start cells in one urban map, not held-out cities. Its target positions/reference trajectory are verifier-only. EB tracks must be grouped as one family when aggregating breadth, rather than three independent benchmark papers.

Reproduction and verification

report.json stores selected/source counts, strata, seed, group partition attempt, catalogue hash, generator hash and exact JSONL hashes. publication.json records package hashes and metadata round-trip validation. Use the Hub commit SHA from the publication receipt or the Hub commit history to pin a run; there are no separate version-named datasets.

The source retrieval script and builder are included as readable provenance; use the source repository to execute them with their dependencies and historical seed files. Keep sources.lock.json fixed. Downloading source archives may take several GiB; put the cache on a disk with room. Source licences apply separately to annotations and media/assets.

The release is a compact local research benchmark, not a replacement for official full benchmark evaluations or evidence that new runtime adapters passed end-to-end evaluation.