Create versions/iclr2027-rebuttal-v1/README.md
Browse files
versions/iclr2027-rebuttal-v1/README.md
ADDED
|
@@ -0,0 +1,56 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# SparseWake supplementary revision artifacts
|
| 2 |
+
|
| 3 |
+
This collection contains two separate, anonymous artifacts for the revised benchmark experiments. The original benchmark release remains fixed.
|
| 4 |
+
|
| 5 |
+
**Licenses:** original code and associated software documentation are licensed under MIT. Original data, observations, targets, metadata, saved predictions, model weights, diagnostic arrays, result tables, protocols and research documentation are licensed under CC BY 4.0. Each archive includes the applicable scope in `LICENSE_STATUS.md`, the full MIT and CC BY 4.0 texts, and `THIRD_PARTY_NOTICES.md`. External dependencies and the separately obtained original core/reproduction assets retain their original licenses and notices.
|
| 6 |
+
|
| 7 |
+
## Archives and coverage
|
| 8 |
+
|
| 9 |
+
| Archive | Contents | Size (bytes) |
|
| 10 |
+
|---|---|---:|
|
| 11 |
+
| `SparseWake_parameter_evaluation_extension_v1.zip` | Additional raw observation histories, targets, sensor coordinates, condition metadata, seed-specific split eligibility, saved predictions and scoring code for observer probe size, relative motion, and source speed under two controlled protocols. | 76,589,128 |
|
| 12 |
+
| `SparseWake_revision_evidence_v1.zip` | Geometry-distance diagnostics, physics-based inverse search evidence, joint source-set models and predictions, frozen-encoder point/Gaussian/mixture heads and predictions, failure-category summaries, protocols and replay/reproduction code. | 192,034,396 |
|
| 13 |
+
|
| 14 |
+
The archives together occupy **268,623,524 bytes** (268.62 MB, 256.18 MiB). SHA-256 checksums are in `SHA256SUMS.txt`; `UPLOAD_MANIFEST.json` describes the public files using relative names only. Each archive also contains its own member-level manifest.
|
| 15 |
+
|
| 16 |
+
### Controlled parameter evaluation
|
| 17 |
+
|
| 18 |
+
The parameter archive includes four families: observer/source length ratios 0.8, 1.0 and 1.2, static and moving histories at five observer speed ratios, and source speed ratios 0.6, 0.8, 1.0, 1.2 and 1.4 under native-ratio and fixed-lateral protocols. There are 19 family-condition entries and 16 distinct physical parameter conditions, with 46,213 stored histories including paired geometries and repeated baselines. These counts are not counts of independent simulations.
|
| 19 |
+
|
| 20 |
+
This is an **evaluation-only** extension of the canonical single-source task. It contains no new training/validation bank. The observer-size condition scales both coordinates of the sensor layout and evaluates probe geometry. It does not simulate coupled fish morphology or observer-induced flow. The existing three reference models use different pose splits, so their pooled pose union is not a universal test set. `reference_split` and `reference_test_eligible` must be used to select each model's own test poses. The archive's `DATA_CARD.json` defines coordinates, array axes and the precise parameter protocols.
|
| 21 |
+
|
| 22 |
+
### Additional model and diagnostic evidence
|
| 23 |
+
|
| 24 |
+
The evidence archive includes six saved multi-hypothesis prediction files and nine trained output heads (point, Gaussian, and mixture) on the three frozen MLP representations. It also includes twelve joint source-set checkpoints, saved source-set predictions, inverse search records, new confirmation observations, and the retained diagnostic sufficient statistics. Machine-specific path metadata in release copies of checkpoints and protocols was replaced with neutral identifiers. Numeric arrays and tensor fields were preserved and checked against the originals.
|
| 25 |
+
|
| 26 |
+
The source-set models use known source count and all-source position supervision. The output heads retain nearest-source supervision on frozen representations. These distinctions remain part of the experiment definitions. The archive provides evidence and metric replay, plus deeper reproduction entry points with stated dependencies. Packaging did not repeat full model training or inverse optimization.
|
| 27 |
+
|
| 28 |
+
## Existing public inputs
|
| 29 |
+
|
| 30 |
+
The frozen base version is [iclr2027-v1.0rc2 at commit b59212263b0ba5762aee76e61f91d36acbefe636](https://huggingface.co/datasets/SparseWake/sparsewake/tree/b59212263b0ba5762aee76e61f91d36acbefe636/versions/iclr2027-v1.0rc2). The following dependencies remain in that existing release and are not duplicated here:
|
| 31 |
+
|
| 32 |
+
| Operation | Required original release files |
|
| 33 |
+
|---|---|
|
| 34 |
+
| Parameter saved-prediction metric replay | None. The extension contains the required arrays. |
|
| 35 |
+
| Revision saved-prediction replay for density heads, stored geometry distances, source-set comparison and failure-category sufficient statistics | None. |
|
| 36 |
+
| Inverse-record metric replay and geometry search checks | `SparseWake_core_1.0rc2.zip`, extracted as `core/`. |
|
| 37 |
+
| Density-head feature preparation and training/evaluation | `SparseWake_core_1.0rc2.zip`, including the frozen MLP encoders and original banks. |
|
| 38 |
+
| Coauthor inverse/source-set deeper reproduction | Both `SparseWake_core_1.0rc2.zip` and `SparseWake_reproduction_assets_1.0rc2.zip`, extracted into sibling `core/` and `reproduction_assets/` directories. |
|
| 39 |
+
|
| 40 |
+
The original ZIPs have no outer container directory. Extract each to the explicitly named destination. The experiment-specific scripts check pinned input hashes. Creating new parameter-generator libraries requires the upstream simulator environment, which is outside these two archives. Replaying the supplied parameter predictions does not require that environment.
|
| 41 |
+
|
| 42 |
+
## Replay from a fresh extraction
|
| 43 |
+
|
| 44 |
+
Extract the two new archives into separate directories named `parameter_extension/` and `revision_evidence/`. Use Python with NumPy. Add h5py for the optional inverse-record replay. Run:
|
| 45 |
+
|
| 46 |
+
```text
|
| 47 |
+
python parameter_extension/score_reference.py --output parameter_metric_replay.json
|
| 48 |
+
python revision_evidence/replay_metrics.py --out revision_metric_replay.json
|
| 49 |
+
python revision_evidence/replay_metrics.py --core PATH_TO_EXTRACTED_CORE --out revision_metric_replay_all.json
|
| 50 |
+
```
|
| 51 |
+
|
| 52 |
+
Write replay outputs outside the extracted artifact directories. `revision_evidence/geometry/independent_check.py`, `multi_hypothesis/REPRODUCE.md` and `coauthor_code/README.md` describe optional deeper checks, environments and entry points. The default replay recomputes stored-output statistics. It does not train networks, rerun the forward simulator or optimize inverse configurations.
|
| 53 |
+
|
| 54 |
+
Independent preparation checks verified ZIP integrity and every archive member hash. Parameter metrics replayed exactly. The combined revision replay checked 674 numerical fields, with maximum absolute difference approximately 1.42e-14. Those checks validate numerical consistency of the retained evidence.
|
| 55 |
+
|
| 56 |
+
These archives contain no DNS pilot, external experimental validation dataset, upstream high-fidelity solver, original private archive, or internal review report.
|