docs: card corrections for v1.2.0 (data unchanged)

README.md

@@ -170,16 +170,16 @@ dataset_info:
     dtype: string
   splits:
   - name: train
-    num_bytes: 4880778432
-    num_examples: 1911
+    num_bytes: 17538156735
+    num_examples: 24104
   - name: validation
-    num_bytes: 407315591
-    num_examples: 196
+    num_bytes: 2907688925
+    num_examples: 4758
   - name: test
-    num_bytes: 1364538389
-    num_examples: 487
-  download_size: 6652632412
-  dataset_size: 6652632412
+    num_bytes: 5186592735
+    num_examples: 5932
+  download_size: 25632438395
+  dataset_size: 25632438395
 - config_name: metadata
   features:
   - name: image_sha256
@@ -908,23 +908,23 @@ Underwater imagery with masks, boxes, points and image labels for coral-reef com
 
 ## Dataset summary
 
-Reef Support Marine Data v1.2.0 keeps the 11 configurations that v1.1.0 and v1.1.1 added to the frozen v1.0 release, additively: image bytes once per `image_sha256`, per-image metadata, point, box, mask and bleaching labels, benthic cover and sites, and the tables that describe the label space (`taxonomy`, `label-map`) and the sources. Every row carries its `image_sha256`, so any label reaches its pixels in one join, and `licence` and attribution stay attached to the pixels. v1.2.0 adds Coralscapes V2 (source id `coralscapes-v2`) to `semseg` and `instances` and adds no config. V2 supersedes V1 for Coralscapes tasks; V1 configs are kept frozen for reproducibility.
+Reef Support Marine Data v1.2.0 keeps the 12 configurations that v1.1.0 (11) and v1.1.1 (`condition-masks`) added to the frozen v1.0 release, additively: image bytes once per `image_sha256`, per-image metadata, point, box, mask and bleaching labels, benthic cover, and the tables that describe the label space (`taxonomy`, `label-map`) and the sources. Every row carries its `image_sha256`, so any label reaches its pixels in one join, and `licence` and attribution stay attached to the pixels. v1.2.0 adds Coralscapes V2 (source id `coralscapes-v2`) to `semseg` and `instances` and adds no config. V2 supersedes V1 for Coralscapes tasks; V1 configs are kept frozen for reproducibility.
 
 | Field | Value |
 |---|---|
-| Configurations | 11: `images`, `metadata`, `points`, `bleaching`, `boxes`, `semseg`, `instances`, `benthic-cover`, `taxonomy`, `label-map`, `sources` |
+| Configurations | 16 (4 from v1.0, 11 added in v1.1.0, `condition-masks` in v1.1.1) |
 | Images (`metadata` rows) | 67,950 |
-| Sources | 15 |
+| Sources | 16 (the `sources` config has 15 rows) |
 | Format | Parquet shards, streamable; images and new masks embedded once |
 | Size of the new configs | 6.9 GB |
-| Version | v1.2.0, built 2026-10-30 |
+| Version | v1.2.0, built 2026-10-30, published 2026-11-01 |
 | Code | [github.com/reefsupport/marine-data](https://github.com/reefsupport/marine-data) |
 
 ## Supported tasks
 
 | Config | Task | Rows | Sources | Notes |
 |---|---|---|---|---|
-| `images` | Pixels for retrieval, pretraining and image classification | 2,594 | `coralscapes-v2`, `noaa-pifsc-bleaching`, `roboflow-coral-bleaching-final-v6i`, `roboflow-coral-bleaching-general-v1-yolov8s`, `roboflow-coral-classification-copy-changed-v13i`, `roboflow-coral-reef-classification-v3i`, `seaview-survey-imagery` | Bytes once per `image_sha256`; images a v1.0 config already embeds are not repeated (`metadata.bytes_config` names the config that holds each image). |
+| `images` | Pixels for retrieval, pretraining and image classification | 34,794 (train 24,104 / validation 4,758 / test 5,932) | `coralscapes-v2`, `noaa-pifsc-bleaching`, `roboflow-coral-bleaching-final-v6i`, `roboflow-coral-bleaching-general-v1-yolov8s`, `roboflow-coral-classification-copy-changed-v13i`, `roboflow-coral-reef-classification-v3i`, `seaview-survey-imagery` | Bytes once per `image_sha256`; images a v1.0 config already embeds are not repeated (`metadata.bytes_config` names the config that holds each image). |
 | `metadata` | Survey and site analytics, joins, retrieval | 67,950 | `coralscapes`, `coralscapes-v2`, `noaa-pifsc-bleaching`, `reef-support-benthic-own`, `reef-support-seaview-labels`, `roboflow-aquarium`, `roboflow-coral-bleaching-final-v6i`, `roboflow-coral-bleaching-general-v1-yolov8s`, `roboflow-coral-classification-copy-changed-v13i`, `roboflow-coral-reef-classification-v3i`, `seaview-survey-imagery`, `suim`, `uiis`, `uiis10k`, `usis10k` | One row per image, v1.0 images included; no pixels. Default config. |
 | `points` | Point-based benthic cover | 858,370 | `seaview-survey-imagery` | One row per labelled point; label-only, no image. |
 | `bleaching` | Image classification (bleaching condition) | 27,542 | `noaa-pifsc-bleaching`, `reef-support-bleaching`, `roboflow-coral-bleaching-final-v6i`, `roboflow-coral-bleaching-general-v1-yolov8s`, `roboflow-coral-classification-copy-changed-v13i`, `roboflow-coral-reef-classification-v3i` | One row per (image, label). |
@@ -934,7 +934,7 @@ Reef Support Marine Data v1.2.0 keeps the 11 configurations that v1.1.0 and v1.1
 | `benthic-cover` | Benthic cover | 3,298 | `coralscapes`, `reef-support-benthic-own` | Derived per image or survey from points and masks. |
 | `taxonomy` | Label space | 3,035 | – | Taxonomy nodes and the edges of the crosswalks in this repo. |
 | `label-map` | Label harmonisation | 401 | `coralscapes`, `coralscapes-v2`, `noaa-pifsc-bleaching`, `reef-support-benthic-own`, `reef-support-seaview-labels`, `roboflow-aquarium`, `roboflow-coral-bleaching-final-v6i`, `roboflow-coral-bleaching-general-v1-yolov8s`, `roboflow-coral-classification-copy-changed-v13i`, `roboflow-coral-reef-classification-v3i`, `seaview-survey-imagery`, `suim`, `uiis`, `uiis10k`, `usis10k` | One row per (source, native label or pixel value) with its taxonomy nodes and match type. |
-| `sources` | Provenance, licence and attribution | 15 | `coralscapes`, `coralscapes-v2`, `noaa-pifsc-bleaching`, `reef-support-benthic-own`, `reef-support-seaview-labels`, `roboflow-aquarium`, `roboflow-coral-bleaching-final-v6i`, `roboflow-coral-bleaching-general-v1-yolov8s`, `roboflow-coral-classification-copy-changed-v13i`, `roboflow-coral-reef-classification-v3i`, `seaview-survey-imagery`, `suim`, `uiis`, `uiis10k`, `usis10k` | One row per source: licence, attribution, rights holder, verification, counts. |
+| `sources` | Provenance, licence and attribution | 15 | `coralscapes`, `coralscapes-v2`, `noaa-pifsc-bleaching`, `reef-support-benthic-own`, `reef-support-seaview-labels`, `roboflow-aquarium`, `roboflow-coral-bleaching-final-v6i`, `roboflow-coral-bleaching-general-v1-yolov8s`, `roboflow-coral-classification-copy-changed-v13i`, `roboflow-coral-reef-classification-v3i`, `seaview-survey-imagery`, `suim`, `uiis`, `uiis10k`, `usis10k` | One row per listed source: licence, attribution, rights holder, verification, counts. `reef-support-bleaching` has no row here (see Sources below). |
 | `coral-masks` | Frozen prior-release config (v1.0) | 5,997 | `coralscapes`, `reef-support-benthic-own`, `reef-support-seaview-labels` | Carried unchanged from v1.0: `train` 4,585, `validation` 569, `test` 843 rows. Not rebuilt by this release; its files stay as published. |
 | `scene-masks` | Frozen prior-release config (v1.0) | 1,598 | `suim` | Carried unchanged from v1.0: `train` 1,181, `validation` 163, `test` 254 rows. Not rebuilt by this release; its files stay as published. |
 | `instance-masks` | Frozen prior-release config (v1.0) | 25,296 | `uiis`, `uiis10k`, `usis10k` | Carried unchanged from v1.0: `train` 19,409, `validation` 2,284, `test` 3,603 rows. Not rebuilt by this release; its files stay as published. |
@@ -992,11 +992,11 @@ Keys: the image key is `image_sha256` (64 lowercase hex, the staged checksum, ne
 
 ## Coralscapes V2 (new in v1.2.0)
 
-**V2 supersedes V1 for Coralscapes tasks; V1 configs are kept frozen for reproducibility.** [Coralscapes V2](https://huggingface.co/datasets/josauder/coralscapesV2) (Apache-2.0, [arXiv:2609.12826](https://arxiv.org/abs/2609.12826)) re-annotates and extends the Coralscapes V1 reef scenes: 2,433 frames of 1024 x 2048 pixels from 41 Red Sea dive sites, labelled with 95 classes instead of 39, including soft corals and black coral. This release pins the upstream revision `8a4567b2fbec14882e09aebff63a7a1ca5e24dca` (card dated 2026-09-15) and keeps the official split: 1,790 train, 170 validation and 473 test frames.
+**V2 supersedes V1 for Coralscapes tasks; V1 configs are kept frozen for reproducibility.** [Coralscapes V2](https://huggingface.co/datasets/josauder/coralscapesV2) (Apache-2.0, [arXiv:2609.12826](https://arxiv.org/abs/2609.12826)) re-annotates and extends the Coralscapes V1 reef scenes: 2,433 frames of 1024 x 2048 pixels from 41 Red Sea dive sites, labelled with 95 classes instead of 39, including soft corals and black coral. This release pins the upstream revision `8a4567b2fbec14882e09aebff63a7a1ca5e24dca` (card dated 2026-09-15). Upstream split 1,790 / 170 / 473. v1.2.0 assigns 1,790 / 170 / 472 (one test frame withheld) and ships 1,786 / 170 / 472 `semseg` rows: four assigned train frames are not shipped; the byte-identical upstream test pair shares one image row, so V2 adds 2,427 images.
 
 ### What changed in `semseg` and `instances`
 
-- Both configs gain rows under the new source id `coralscapes-v2` (2,427 images in this release); the `sources` config and the Splits table carry the counts per source, config and split. No config is added or renamed, and no existing file, row, id or path changes.
+- Both configs gain rows under the new source id `coralscapes-v2` (2,427 images in this release); the `sources` config and the Splits table carry the counts per source, config and split. No config is removed or renamed. The four v1.0 configs, `condition-masks`, `benthic-cover` and every earlier `images` shard are byte-identical to v1.1.1. v1.2.0 rewrites 91 Parquet files of the other configs in place (new rows, the 9 moved split groups, `taxonomy_version` 3.1.0); no split loses rows except the one UIIS10K image moved to `train`.
 - `semseg` holds the V2 semantic masks as `uint8` PNGs with the upstream class ids 1 to 95. Pixel `0` is the upstream "unlabelled" value and is not counted as a class; `label-map` records the ignore semantics per source.
 - `instances` holds the V2 instance masks as `uint16` PNGs, unchanged from upstream: a pixel below 2000 is the class id of a stuff class, a pixel of 2000 or more is `class_id * 2000 + instance_index`, and fish is class 25. The upstream card reports 65,738 fish instance masks.
 - `mask_ref` is the object key in the public bucket for the sources built by the unified masks table (in v1.2.0: Coralscapes V2, `sources/<source_id>/<version>/labels/...`), and the path relative to the source's staged tree `sources/<source_id>/<version>/` for the earlier sources.
@@ -1011,13 +1011,13 @@ Every V2 class resolves into `rs-benthic-v1` through the crosswalk `coralscapes-
 
 ### Splits: V1 and V2 twins
 
-2,074 V2 frames are V1 frames: a V1 file name is the V2 stem (`site<N>_<frame>_<timestamp>`) plus `_leftImg8bit.png`. V2 re-encoded the images and re-annotated the masks, so the bytes and the pixels of a twin differ from V1 and the sha256 keys that carry the v1.0 splits forward cannot link them. The split map therefore links twins by stem and treats each pair as one group. The dedup and splits gates count a pair in one split as one link group and fail a pair that sits in two splits. For all 2,074 twins the release split of the V1 frame equals the official V2 split, so no frame moves. V2 adds 359 frames (273 train, 4 validation, 82 test) that keep their official V2 split; the V2 test and validation frames are pinned as benchmark splits (472 of the 473 test frames and all 170 validation frames; the remaining test frame is withheld, see the split map below). 1 V1 frame, a V1 test frame, has no V2 twin and exists only in the frozen V1 configs. V2 covers 41 dive sites (V1: 35; the six new ones are site36 to site41). The 55 V2-only frames of site30 are V2 train frames while the frames of site30 shared with V1 are validation frames; the official split is kept and the gates re-check that boundary at build time. V1 and V2 counts are not additive: summing them counts the 2,074 shared frames twice.
+2,074 V2 frames are V1 frames: a V1 file name is the V2 stem (`site<N>_<frame>_<timestamp>`) plus `_leftImg8bit.png`. V2 re-encoded the images and re-annotated the masks, so the bytes and the pixels of a twin differ from V1 and the sha256 keys that carry the v1.0 splits forward cannot link them. The split map therefore links twins by stem and treats each pair as one group. The dedup and splits gates count a pair in one split as one link group and fail a pair that sits in two splits. For all 2,074 twins the release split of the V1 frame equals the official V2 split, so no twin changes split. V2 adds 359 frames (273 train, 4 validation, 82 test) that keep their official V2 split; the V2 test and validation frames are pinned as benchmark splits (472 of the 473 test frames and all 170 validation frames; the remaining test frame is withheld, see the split map below). 1 V1 frame, a V1 test frame, has no V2 twin and exists only in the frozen V1 configs. V2 covers 41 dive sites (V1: 35; the six new ones are site36 to site41). The 55 V2-only frames of site30 are V2 train frames while the frames of site30 shared with V1 are validation frames; the official split is kept and the gates re-check that boundary at build time. V1 and V2 counts are not additive: summing them counts the 2,074 shared frames twice.
 
 ### Split map of v1.2.0
 
-**Inherited.** The split map inherits the v1.1.0 map (`inherits` in `SPLIT_MAP.json`): its 35,825 groups carry over, 4,382 of them kept verbatim, and the splits that v1.1.1 published per image are the prior for the rest. Re-resolving them changed the split of 9 inherited groups and of no other: 8 are near-duplicate groups from the Roboflow bleaching exports that no published v1.1.x table references, and 1 is a UIIS10K image that v1.1.1 published in two splits (train in `fish-boxes` and `instance-masks`, validation in `metadata`); v1.2.0 puts it in train in every table. No benchmark pin overrides an inherited split (`pins.v1_continuity_overridden`: 0 images, 0 groups).
+**Inherited.** The split map inherits the v1.1.0 map (`inherits` in `SPLIT_MAP.json`, which is not on the Hub: it is frozen in the GitHub repo at `releases/marine-data/v1.2.0/SPLIT_MAP.json`): its 35,825 groups carry over, 4,382 of them kept verbatim, and the splits that v1.1.1 published per image are the prior for the rest. Re-resolving them changed the split of 9 inherited groups and of no other: 8 are near-duplicate groups from the Roboflow bleaching exports that no published v1.1.x table references, and 1 is a UIIS10K image that v1.1.1 published in two splits (train in `fish-boxes` and `instance-masks`, validation in `metadata`); v1.2.0 puts it in train in every table. No benchmark pin overrides an inherited split (`pins.v1_continuity_overridden`: 0 images, 0 groups).
 
-**Coralscapes V2.** 2,432 of the 2,433 V2 frames are assigned: 1,790 to train, 170 to validation and 472 to test (the official V2 split has 473 test frames; the one missing is withheld, below). The V2 benchmark pins hold 472 test and 170 validation images, counted by digest, so the byte-identical test pair counts once; the V1 pins are unchanged at 391 test and 166 validation images. Stem links join the V1 and V2 twins: 2,074 twins, 359 V2-only frames, 1 V1-only frame, 0 ambiguous stems and 0 conflicts (a conflict is a twin whose upstream V2 split differs from the split v1.0 published).
+**Coralscapes V2.** 2,432 of the 2,433 V2 frames are assigned: 1,790 to train, 170 to validation and 472 to test (the official V2 split has 473 test frames; the one missing is withheld, below). v1.2.0 ships 1,786 / 170 / 472 `semseg` rows: four assigned train frames are not shipped; the byte-identical upstream test pair shares one image row, so V2 adds 2,427 images. The V2 benchmark pins hold 472 test and 170 validation images, counted by digest, so the byte-identical test pair counts once; the V1 pins are unchanged at 391 test and 166 validation images. Stem links join the V1 and V2 twins: 2,074 twins, 359 V2-only frames, 1 V1-only frame, 0 ambiguous stems and 0 conflicts (a conflict is a twin whose upstream V2 split differs from the split v1.0 published).
 
 **Withheld and left out.** One group is withheld: a V2 benchmark frame inside the near-duplicate band of train images of a prior release. The band says never-eval and the pin says eval, so the frame ships in no table of any config. The split map also leaves out 18 images that sit in the prior-release near-duplicate band (dHash distance 8 or less) of prior-release train and evaluation images, on the side their near-duplicate group did not keep, and 34 images that each collide on the DCT pHash-64 with an image of another split (one image of the pair goes, never an image that an earlier release published); these 52 images ship in no split of any config.
 
@@ -1029,7 +1029,7 @@ Two V2 frames, `site23_000097_017682` and `site23_000152_017682`, have identical
 
 ### What stays frozen
 
-The four v1.0 configs (`coral-masks`, `scene-masks`, `instance-masks`, `fish-boxes`) and every v1.1.0 and v1.1.1 config keep their files, rows, ids and paths; v1.2.0 only adds files. Coralscapes V1 (`coralscapes`) stays where it is, with its V1 splits, so earlier tags reproduce exactly. Pin a tag with `revision` to read an earlier release. Downstream model pins are unchanged.
+No config is removed or renamed. The four v1.0 configs (`coral-masks`, `scene-masks`, `instance-masks`, `fish-boxes`), `condition-masks`, `benthic-cover` and every earlier `images` shard are byte-identical to v1.1.1. v1.2.0 rewrites 91 Parquet files of the other configs in place (new rows, the 9 moved split groups, `taxonomy_version` 3.1.0); no split loses rows except the one UIIS10K image moved to `train`. Coralscapes V1 (`coralscapes`) stays where it is, with its V1 splits, so earlier tags reproduce exactly. Pin a tag with `revision` to read an earlier release. Downstream model pins are unchanged.
 
 ### Citing Coralscapes V2
 
@@ -1105,7 +1105,7 @@ Taxonomy nodes that more than one source reaches. Cross-source training should u
 | `SP` | 2 | `coralscapes`, `coralscapes-v2` |
 | `TL` | 2 | `coralscapes`, `coralscapes-v2` |
 
-46 node(s) are reached by a single source.
+46 nodes are reached by a single source.
 
 ### Mapped share per source
 
@@ -1145,21 +1145,21 @@ A crosswalk makes labels comparable, not identical. A `coarsened` edge loses det
 
 ## Splits
 
-The card publishes the **measured** split shares of every config, never a nominal ratio: upstream splits dominate several sources, so the v1.0 shares are not 80/10/10 or 70/15/15. New rows target 70/15/15 with the split names `train`, `validation`, `test` and `test_ood`. Precedence, first match wins: (1) an image that matches a v1.0 image by sha256, or within Hamming distance 4 of its dhash, keeps that v1.0 split unless an upstream benchmark manifest pins it (see below), and an image within 5-8 of a v1.0 image is never used across the train/eval boundary; (2) the out-of-distribution holdout rules (deepfish, temperate-australasia, mediterranean, a data-driven tropical province, depth of 800 m or more, platform AUV, year 2025 or later) go to `test_ood`, which is never mixed into `test`; (3) the upstream split where defined; (4) a group hash on survey, site or quadrat, else the near-duplicate group. Images of one group share a split. Group keys in `SPLIT_MAP.json` are namespaced `<source>/<id>`, so equal ids in two sources are two groups.
+The card publishes the **measured** split shares of every config, never a nominal ratio: upstream splits dominate several sources, so the v1.0 shares are not 80/10/10 or 70/15/15. New rows target 70/15/15 with the split names `train`, `validation`, `test` and `test_ood`. Precedence, first match wins: (1) an image that matches a v1.0 image by sha256, or within Hamming distance 4 of its dhash, keeps that v1.0 split unless an upstream benchmark manifest pins it (see below), and an image within 5-8 of a v1.0 image is never used across the train/eval boundary; (2) the out-of-distribution holdout rules (deepfish, temperate-australasia, mediterranean, a data-driven tropical province, depth of 800 m or more, platform AUV, year 2025 or later) go to `test_ood`, which is never mixed into `test`; (3) the upstream split where defined; (4) a group hash on survey, site or quadrat, else the near-duplicate group. Images of one group share a split. Group keys in `SPLIT_MAP.json` (GitHub repo, `releases/marine-data/v1.2.0/SPLIT_MAP.json`) are namespaced `<source>/<id>`, so equal ids in two sources are two groups.
 
 Version 1.1.0 inherits the v1.0 splits by `image_sha256` from the published v1.0 tree, with one exception: six benchmark sources (`coralscapes`, `suim`, `uiis`, `uiis10k`, `usis10k` and one benchmark not distributed in the open release) are pinned to their upstream manifests. A pin routes the manifest's evaluation split to `test` and a second held-out split, where the source has one, to `validation`. A pin wins over an inherited v1.0 split; the out-of-distribution holdout and any rule that keeps an image out of evaluation win over a pin. The pins moved images that were in v1.0 `train` into the v1.1.0 `validation` and `test` splits and resolved near-duplicate groups that straddle v1.0 splits, while the frozen v1.0 configs keep their v1.0 splits. A group that cannot be assigned consistently is withheld from every table and listed under `withheld` in `SPLIT_MAP.json`.
 
 Because several of these sources carry large upstream test sets, the overall split is not the 70/15/15 target: the test share is well above 15 %, and `boxes` and `instances` are the most affected tables.
 
-**Evaluation caveat.** A model trained on the v1.0 configs must not be evaluated on the v1.1.0 `validation` or `test` splits of these six sources, because some of their images were in `train` in v1.0.
+**Evaluation caveat.** A model trained on the v1.0 configs must not be evaluated on the v1.1.0 and later `validation` or `test` splits of these six sources, because some of their images were in `train` in v1.0.
 
-**Left out and overridden (v1.1.0).** No group was withheld. The split map also leaves out 215 images that sit in the prior-release near-duplicate band (dHash distance 5 to 8) of v1.0 train and evaluation images, on the side their near-duplicate group did not keep, and 51 images that each collide on the DCT pHash-64 with an image of another split (one image of the pair goes, never a v1.0 image); these 266 images ship in no split of any config. Where a benchmark pin and v1.0 continuity disagree the pin wins, and it resolves the near-duplicate groups that straddle v1.0 splits: the pins overrode the v1.0 split of 4,544 images (`pins.v1_continuity_overridden` in `SPLIT_MAP.json`).
+**Left out and overridden (v1.1.0; the v1.2.0 figures are under "Split map of v1.2.0" above).** No group was withheld. The split map also leaves out 215 images that sit in the prior-release near-duplicate band (dHash distance 5 to 8) of v1.0 train and evaluation images, on the side their near-duplicate group did not keep, and 51 images that each collide on the DCT pHash-64 with an image of another split (one image of the pair goes, never a v1.0 image); these 266 images ship in no split of any config. Where a benchmark pin and v1.0 continuity disagree the pin wins, and it resolves the near-duplicate groups that straddle v1.0 splits: the pins overrode the v1.0 split of 4,544 images (`pins.v1_continuity_overridden` in `SPLIT_MAP.json`).
 
-**Ratios.** The 70/15/15 target is checked per source family, over distinct images, not per config. The NOAA bleaching images, the Roboflow Coral Universe forks and the Seaview sources measure 70.8 / 14.6 / 14.6 %, 68.5 / 15.7 / 15.8 % and 67.6 / 12.6 / 19.8 % (train / validation / test); pinned benchmark sources and strata under 100 images (`roboflow-aquarium`) are left out of the check. Row shares of a config therefore deviate, as the table shows: `points` rows are 65.2 / 12.7 / 22.1 % (its 11,356 distinct images 67.1 / 12.8 / 20.2 %) and `semseg` rows 74.1 / 8.9 / 16.9 %. `noaa-pifsc-bleaching` is assigned by whole site: its 236 sites went 42 to train, 97 to validation and 97 to test, so its image shares are not its site shares.
+**Ratios.** The 70/15/15 target is checked per source family, over distinct images, not per config. The NOAA bleaching images, the Roboflow Coral Universe forks and the Seaview sources measure 70.8 / 14.6 / 14.6 %, 68.5 / 15.7 / 15.8 % and 67.6 / 12.6 / 19.8 % (train / validation / test); pinned benchmark sources and strata under 100 images (`roboflow-aquarium`) are left out of the check. Row shares of a config therefore deviate, as the table shows: `points` rows are 65.2 / 12.7 / 22.1 % (its 11,356 distinct images 67.1 / 12.8 / 20.2 %) and `semseg` rows 73.9 / 8.1 / 18.0 %. `noaa-pifsc-bleaching` is assigned by whole site: its 236 sites went 42 to train, 97 to validation and 97 to test, so its image shares are not its site shares.
 
 | Config | train | validation | test | test_ood | Total |
 |---|---|---|---|---|---|
-| `images` | 1,911 (73.7 %) | 196 (7.6 %) | 487 (18.8 %) | – | 2,594 |
+| `images` | 24,104 (69.3 %) | 4,758 (13.7 %) | 5,932 (17.0 %) | – | 34,794 |
 | `metadata` | 42,095 (61.9 %) | 8,291 (12.2 %) | 17,564 (25.8 %) | – | 67,950 |
 | `points` | 559,540 (65.2 %) | 108,730 (12.7 %) | 190,100 (22.1 %) | – | 858,370 |
 | `bleaching` | 19,766 (71.8 %) | 3,865 (14.0 %) | 3,911 (14.2 %) | – | 27,542 |
@@ -1181,6 +1181,7 @@ Because several of these sources carry large upstream test sets, the overall spl
 | `coralscapes-v2` | 2.0 | – | dataset-card 2026-10-29 | https://josauder.github.io/coralscapesv2/ | 2,427 | images,instances,metadata,semseg |
 | `noaa-pifsc-bleaching` | 1-image-labels | – | dataset-card 2026-08-17 | – | 10,415 | bleaching,images,metadata |
 | `reef-support-benthic-own` | 2026-10-06 | – | licence-file 2026-08-17 | – | 1,226 | benthic-cover,metadata,semseg |
+| `reef-support-bleaching` | 2026-09-24 | – | licence-file 2026-08-17 | – | 635 | bleaching,condition-masks |
 | `reef-support-seaview-labels` | 2026-10-06 | – | licence-file 2026-10-06 | – | 2,707 | metadata |
 | `roboflow-aquarium` | zip-5d30fed3dd5a | – | licence-file 2026-10-06 | https://public.roboflow.ai/object-detection/aquarium | 638 | boxes,metadata |
 | `roboflow-coral-bleaching-final-v6i` | v6i-image-labels-r2 | – | licence-file 2026-09-24 | – | 2,532 | bleaching,images,metadata |
@@ -1193,22 +1194,24 @@ Because several of these sources carry large upstream test sets, the overall spl
 | `uiis10k` | rev-5c4316368a8e | – | dataset-card 2026-10-06 | https://huggingface.co/datasets/LiamLian0727/UIIS10K | 10,042 | boxes,instances,metadata |
 | `usis10k` | rev-b10f41ab819b | – | dataset-card 2026-10-06 | https://huggingface.co/datasets/LiamLian0727/USIS10K | 10,611 | boxes,instances,metadata |
 
+`coralscapes-v2` counts 2,427 images because the byte-identical upstream test pair shares one image row (its `semseg` rows number 2,428). `reef-support-bleaching` has no row in the `sources` config; its 635 images are the Tayrona images of `condition-masks`, and 634 of them also carry `bleaching` rows.
+
 ### Build and gates
 
-Built with the [marine-data](https://github.com/reefsupport/marine-data) registry; every source is admitted only after its licence is recorded, and the licence gate raises, never warns. The release is pinned by `provenance/release.lock.yaml`. Gate results are published as `provenance/GATES.json`. Near-duplicate images inside one split that come from different Roboflow Universe forks of one image set, or from the Seaview survey imagery and our labelled copies of its photographs, are disclosed, not removed: the dedup gate counts them as `near_pairs_across_sources_in_family` in `provenance/GATES.json`, and a near-duplicate across splits still fails it. In v1.1.0, 5,201 such pairs are disclosed. A further 6 pairs from different sources fall within the pHash-64 threshold (8 bits) but are kept: the pHash-256 of the two images differs by more than 32 bits and both carry at least 4 bits of grey entropy, so the gate waives them as look-alikes and counts them as `near_pairs_across_sources_unconfirmed`.
+Built with the [marine-data](https://github.com/reefsupport/marine-data) registry; every source is admitted only after its licence is recorded, and the licence gate raises, never warns. The release is pinned by `provenance/release.lock.yaml`. Gate results are published as `provenance/GATES.json`. Near-duplicate images inside one split that come from different Roboflow Universe forks of one image set, or from the Seaview survey imagery and our labelled copies of its photographs, are disclosed, not removed: the dedup gate counts them as `near_pairs_across_sources_in_family` in `provenance/GATES.json`, and a near-duplicate across splits still fails it. In v1.1.0, 5,201 such pairs are disclosed. A further 6 pairs from different sources fall within the pHash-64 threshold (8 bits) but are kept: the pHash-256 of the two images differs by more than 32 bits and both carry at least 4 bits of grey entropy, so the gate waives them as look-alikes and counts them as `near_pairs_across_sources_unconfirmed`. In v1.2.0 the gates ran on the 2,594 images added in this release; the earlier 32,200 were gated in v1.1.x.
 
 **v1.2.0 gates.** 16 gates ran on the build host. The splits gate checks the 2074 V1/V2 stem-linked twin pairs and finds none split apart. The dedup gate treats a near-duplicate pair of linked twins as one group rather than a leak and reports 0 such pairs among the images stored in this release; V1 pixels stay in the earlier shards and are not re-scanned, so that count covers only pairs shipped here. The numbers are in `provenance/GATES.json`.
 
 ## Personal and sensitive information
 
-The images are underwater and reef photographs. Divers, and rarely faces, can appear. The release gates include a privacy pass (face detections at a score of 0.85 or more are excluded, 0.60 to 0.85 are flagged) and a check for EXIF GPS; their results are published in `provenance/GATES.json`. Images the detector flags are reviewed on contact sheets: an image with an identifiable person is withheld from every table, and a detection that proves to be a false positive is recorded in `registry/privacy/review.yaml` in the code repository. No image with an identifiable person is published knowingly; one Roboflow fork was excluded from this release for that reason (see Excluded sources). A gate and a review lower the risk; they do not remove it. The EXIF GPS of `seaview-survey-imagery` images is retained by design, because that source publishes per-image coordinates. `metadata` carries position at the precision the source provides (`coord_precision_m`) and no source is currently flagged location-sensitive in the registry. In v1.1.0 the privacy gate scanned 32,200 images: 10 flagged images were reviewed by hand and cleared, none was withheld and no unreviewed detection remains. Divers are an annotated class, kept on purpose: `coralscapes`, `uiis`, `uiis10k` and `usis10k` label them (`human` and `human divers`, 7 `label-map` rows). No face or person identification is intended or supported. To drop the class, filter out the rows whose `taxon_node_id` ends in `NT_DIVER` (`label-map` lists the labels that map to it) and send its value to `255` (ignore) in the masks.
+The images are underwater and reef photographs. Divers, and rarely faces, can appear. The release gates include a privacy pass (face detections at a score of 0.85 or more are excluded, 0.60 to 0.85 are flagged) and a check for EXIF GPS; their results are published in `provenance/GATES.json`. Images the detector flags are reviewed on contact sheets: an image with an identifiable person is withheld from every table, and a detection that proves to be a false positive is recorded in `registry/privacy/review.yaml` in the code repository. No image with an identifiable person is published knowingly; one Roboflow fork was excluded from this release for that reason (see Excluded sources). A gate and a review lower the risk; they do not remove it. The EXIF GPS of `seaview-survey-imagery` images is retained by design, because that source publishes per-image coordinates. `metadata` carries position at the precision the source provides (`coord_precision_m`) and no source is currently flagged location-sensitive in the registry. In v1.1.0 the privacy gate scanned 32,200 images: 10 flagged images were reviewed by hand and cleared, none was withheld and no unreviewed detection remains. In v1.2.0 the gates ran on the 2,594 images added in this release; the earlier 32,200 were gated in v1.1.x. Divers are an annotated class, kept on purpose: `coralscapes`, `uiis`, `uiis10k` and `usis10k` label them (`human` and `human divers`, 7 `label-map` rows). No face or person identification is intended or supported. To drop the class, filter out the rows whose `taxon_node_id` ends in `NT_DIVER` (`label-map` lists the labels that map to it) and send its value to `255` (ignore) in the masks.
 
 ## Limitations and known defects
 
 - Coralscapes V1 (`coralscapes`) has no soft-coral or octocoral class. Coralscapes V2 (`coralscapes-v2`) has soft-coral and black-coral classes, so use V2 for Coralscapes tasks. `label-map` records the classes each source cannot contain, so a model trained on V1 alone cannot predict them and cover for them is undercounted. A V1 frame and its V2 twin are the same frame: the counts of the two sources are not additive (2,074 frames are in both).
 - Absence is not a negative. For the reef-support and Seaview label sources `255` means unannotated, not background; Coralscapes uses its own ignore id. `label-map` carries the per-source ignore semantics.
 - Cross-source benthic training should use the coarse training views, or mask the classes a source cannot contain: the label overlap table shows how few classes more than one source reaches.
-- Labels without a taxonomy mapping are kept (`mapped = false`) and ignored in the training views. A source whose mapped share is below the 95 % floor is listed in the sources table with its share.
+- Labels without a taxonomy mapping are kept (`mapped = false`) and ignored in the training views. A source whose mapped share is below the 95 % floor is listed in the sources table.
 - Until the declared bridges from `rs-taxa-v1` to `rs-benthic-v1` and `rs-fauna-v1` exist, rows labelled by an identity node carry a null `rs_benthic_code`.
 - A small number of identical images carry conflicting labels; every party has `label_status = conflict` and all rows are kept. The sampled label-quality audit has not been run.
 - Confidence is present only where the upstream source supplies it; forks and crowd sources have none. It is never invented.
@@ -1226,10 +1229,11 @@ The data are not under one licence. Every row records its `licence`, `annotation
 | Source | Image licence | Annotation licence | Attribution | Class |
 |---|---|---|---|---|
 | `coralscapes` | [Apache-2.0](https://www.apache.org/licenses/LICENSE-2.0) | Apache-2.0 | Coralscapes (Apache-2.0) | open |
-| `coralscapes-v2` | [Apache-2.0](https://www.apache.org/licenses/LICENSE-2.0) | Apache-2.0 | Coralscapes V2 (Apache-2.0) | open |
+| `coralscapes-v2` | [Apache-2.0](https://www.apache.org/licenses/LICENSE-2.0) | Apache-2.0 | Coralscapes V2 (Apache-2.0); see "Citing Coralscapes V2" | open |
 | `noaa-pifsc-bleaching` | US-GOV-PD | US-GOV-PD | NOAA PIFSC-ESD coral bleaching dataset (US-GOV-PD) | open |
 | `reef-support-benthic-own` | [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) | CC-BY-4.0 | Reef Support benthic instance masks — own imagery (Colombia) (CC-BY-4.0) | open |
-| `reef-support-seaview-labels` | [CC BY 3.0](https://creativecommons.org/licenses/by/3.0/) | CC-BY-3.0 | Reef Support annotations on XL Catlin Seaview imagery (CC-BY-3.0) | open |
+| `reef-support-bleaching` | [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) | CC-BY-4.0 | Reef Support bleaching masks (Tayrona) (CC-BY-4.0) | open |
+| `reef-support-seaview-labels` | [CC BY 3.0](https://creativecommons.org/licenses/by/3.0/) | CC-BY-4.0 | Reef Support annotations on XL Catlin Seaview imagery: CC-BY-3.0 (images); CC-BY-4.0 (masks, labels) | open |
 | `roboflow-aquarium` | [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) | CC-BY-4.0 | Roboflow Aquarium Combined (CC-BY-4.0) | open |
 | `roboflow-coral-bleaching-final-v6i` | [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) | CC-BY-4.0 | Coral Bleaching Final (Roboflow v6i) (CC-BY-4.0) | open |
 | `roboflow-coral-bleaching-general-v1-yolov8s` | [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) | CC-BY-4.0 | Coral Bleaching_General (Roboflow v1, YOLOv8s) (CC-BY-4.0) | open |
@@ -1256,7 +1260,7 @@ Every registered source that is not in this release falls under one of the reaso
 - **sonar sources**: side-scan and forward-looking sonar imagery is a different sensing modality from the camera images of this release; none is included
 - **open-licence sources not yet staged**: registered sources with an open licence that were not in the v1.1.0 fetch list (for example `noaa-pifsc-esa-coral-icra`, `marineevt` and `salmon-cage`) are deferred to a later release
 - **non-image sources**: text-only sources such as `oceaninstruct` hold no images
-- **One non-commercial source**: licensed for non-commercial use only; it ships in the separate non-commercial flavour of the release, not in this repository. No other non-commercial source is part of v1.1.0
+- **One non-commercial source**: licensed for non-commercial use only; it ships in the separate non-commercial flavour of the release, not in this repository. No other non-commercial source is part of v1.2.0
 
 ## Takedown
 
@@ -1266,28 +1270,28 @@ Rights holders and data subjects: open an issue at https://github.com/reefsuppor
 
 | Version | Date | Notes |
 |---|---|---|
-| v1.0.0 | 2026-10-07 | coral-masks, scene-masks, instance-masks and fish-boxes: 8 sources, 58,803 rows. Frozen: files, rows and paths are unchanged in v1.1.0. |
-| v1.1.0 | 2026-10-17 | Additive: 11 new configurations. |
-| v1.1.1 | 2026-10-24 | Additive: one new configuration, `condition-masks` (634 Tayrona bleaching rasters, `train`). Every v1.1.0 file is unchanged. |
-| v1.2.0 | 2026-10-30 | Additive: Coralscapes V2 added to `semseg` and `instances`; every v1.0, v1.1.0 and v1.1.1 file, row, id and path is unchanged. |
+| v1.0 | 2026-10-07 | coral-masks, scene-masks, instance-masks and fish-boxes: 8 sources, 58,803 rows. Frozen: files, rows and paths are unchanged in v1.1.0. |
+| v1.1.0 | 2026-10-18 | Additive: 11 new configurations. |
+| v1.1.1 | 2026-10-24 | Additive: one new configuration, `condition-masks` (635 Tayrona bleaching rasters, `train`). Every v1.1.0 file is unchanged. |
+| v1.2.0 | 2026-11-01 | Coralscapes V2 added to `semseg` and `instances`, with new rows in `images`, `metadata`, `bleaching`, `points`, `label-map`, `sources` and `taxonomy`. No config is removed or renamed. The four v1.0 configs, `condition-masks`, `benthic-cover` and every earlier `images` shard are byte-identical to v1.1.1. v1.2.0 rewrites 91 Parquet files of the other configs in place (new rows, the 9 moved split groups, `taxonomy_version` 3.1.0); no split loses rows except the one UIIS10K image moved to `train`. |
 
 **v1.2.0**
 
-- **Added (additive; no v1.0, v1.1.0 or v1.1.1 file, row, id or path changed):** Coralscapes V2 (source id `coralscapes-v2`, 2,433 frames, 95 classes) in `semseg` and `instances`. No config is added or renamed.
+- **Added:** Coralscapes V2 (source id `coralscapes-v2`, 95 classes; 2,433 upstream frames, 2,427 images in this release) in `semseg` and `instances`. No config is removed or renamed. The four v1.0 configs, `condition-masks`, `benthic-cover` and every earlier `images` shard are byte-identical to v1.1.1.
 - **Superseded, not removed:** V2 supersedes V1 for Coralscapes tasks; V1 configs are kept frozen for reproducibility.
 - **Label space:** taxonomy 3.1.0 adds three `rs-benthic-v1` nodes (`CN_ANTIPATHARIA`, `SC_ALCYONIIDAE`, `SC_NEPHTHEIDAE`) and the 95 `coralscapes-95` edges; nothing in 3.0.0 changes. The `coral-binary` view counts black coral (`CN_ANTIPATHARIA`) as coral.
-- **Splits:** the 2,074 frames V1 and V2 share are linked by stem, so a twin pair is one group; the official V2 split is kept (train 1,790 / validation 170 / test 473) and no frame moves.
-- **Changed:** this card (text only).
+- **Splits:** the 2,074 frames V1 and V2 share are linked by stem, so a twin pair is one group; no twin changes split. Upstream split 1,790 / 170 / 473. v1.2.0 assigns 1,790 / 170 / 472 (one test frame withheld) and ships 1,786 / 170 / 472 `semseg` rows: four assigned train frames are not shipped; the byte-identical upstream test pair shares one image row, so V2 adds 2,427 images.
+- **Changed:** 91 existing Parquet files are rewritten in place (new rows, the 9 moved split groups, `taxonomy_version` 3.1.0), in `metadata` (+2,594 rows), `bleaching` (+199), `points` (+50), `semseg` (+2,428), `instances` (+63,993), `label-map` (+95), `sources` (+1), `taxonomy` (+98) and `boxes` (one UIIS10K image moves from `validation` to `train`; 93,958 rows before and after). `images` gains 28 shards and 2,594 images (2,427 from Coralscapes V2, 167 from six earlier sources not shipped before); its earlier shards are untouched. No split loses rows except the one UIIS10K image moved to `train`. `LICENSE`, `NOTICE`, `CITATION.cff`, `croissant.json`, `provenance/` and this card are rewritten as well.
 
 **v1.1.1**
 
-- Additive: one new configuration, `condition-masks` (634 Tayrona bleaching rasters, `train`). Every v1.1.0 file is unchanged.
+- Additive: one new configuration, `condition-masks` (635 Tayrona bleaching rasters, `train`). Every v1.1.0 file is unchanged.
 
 **v1.1.0**
 
-- **Added (additive; no v1.0 file, row, path or `configs:` entry changed):** the configs `images`, `metadata`, `points`, `bleaching`, `boxes`, `semseg`, `instances`, `benthic-cover`, `sites`, `label-map` and `sources`, and `taxonomy` in the open repo; `LICENSE`, `NOTICE`, `CITATION.cff`, `croissant.json` and `provenance/`.
+- **Added (additive; no v1.0 file, row, path or `configs:` entry changed):** the configs `images`, `metadata`, `points`, `bleaching`, `boxes`, `semseg`, `instances`, `benthic-cover`, `label-map` and `sources`, and `taxonomy` in the open repo; `LICENSE`, `NOTICE`, `CITATION.cff`, `croissant.json` and `provenance/`.
 - **Unchanged:** every v1.0 config, with its files, rows and paths; the release is additive.
-- **Task layers:** the v2 task layers are new tables on the `image_sha256` key: `metadata`, `points`, `bleaching`, `boxes`, `semseg` and `instances` (`benthic-cover` and `sites` are derived from them).
+- **Task layers:** the v2 task layers are new tables on the `image_sha256` key: `metadata`, `points`, `bleaching`, `boxes`, `semseg` and `instances` (`benthic-cover` is derived from them).
 - **Changed:** this card (text only; the v1.0 `configs:` entries are carried verbatim). `dataset_info` is written for the new configs only.
 - **Label space:** taxonomy 3.0.0. Every new table row carries `taxonomy_version`; the Coralscapes class "unknown hard substrate" is unmapped in the new tables; three wrong AphiaIDs were corrected (205902, 206485, 758259). v1.0 `class_map` columns are untouched.
 - **Splits:** a split map that inherits the v1.0 splits by `image_sha256` (v1.0 continuity); six benchmark sources are pinned to their upstream manifests instead, which moved some images out of v1.0 `train` into `validation` and `test`; the frozen v1.0 configs keep their v1.0 splits (see Splits).
@@ -1315,7 +1319,7 @@ load_dataset("reefsupport/marine-data", "metadata", split="train", revision="v1.
 Please also cite the sources you use:
 
 - `coralscapes`: Sauder et al., The Coralscapes Dataset, ICCV 2025 Marine Vision Workshop
-- `coralscapes-v2`: Sauder, Ruckli, Strodomskytė, Tuia, Banc-Prandi et al., CoralscapesV2: Panoptic and Fine-Grained Visual Scene Understanding in Coral Reefs (dataset release, 2026), arXiv:2609.12826. Dataset card josauder/coralscapesV2 at revision 8a4567b2fbec14882e09aebff63a7a1ca5e24dca (card dated 2026-09-15). The card's own tags and BibTeX still point at the V1 paper (arXiv:2503.20000); cite 2609.12826 for V2.
+- `coralscapes-v2`: Sauder, Ruckli, Strodomskytė, Tuia, Banc-Prandi et al., CoralscapesV2: Panoptic and Fine-Grained Visual Scene Understanding in Coral Reefs (dataset release, 2026), arXiv:2609.12826. Dataset card josauder/coralscapesV2 at revision 8a4567b2fbec14882e09aebff63a7a1ca5e24dca (card dated 2026-09-15). Cite Coralscapes V2 with the BibTeX from the upstream dataset card (arXiv:2609.12826) and Coralscapes V1 with arXiv:2503.20000.
 - `noaa-pifsc-bleaching`: Pacific Islands Fisheries Science Center (2025). Ecosystem Sciences Division (ESD). NOAA-ESD-CORAL-Bleaching-Dataset v1. https://huggingface.co/datasets/NMFS-OSI/NOAA-PIFSC-ESD-CORAL-Bleaching-Dataset
 - `reef-support-seaview-labels`: González-Rivero et al. (2019). Seaview Survey Photo-quadrat and Image Classification Dataset. University of Queensland. doi:10.14264/UQL.2019.930 Masks and labels: Reef Support annotations (CC-BY-4.0).
 - `roboflow-aquarium`: Roboflow 2020, Aquarium Combined (public.roboflow.ai/object-detection/aquarium)