Minimal card for now

README.md

@@ -2,14 +2,6 @@
 license: other
 license_name: per-source-terms
 pretty_name: Instruct-181K
-task_categories:
-- visual-question-answering
-- any-to-any
-tags:
-- thinkmorph-ssl
-- primer
-- interleaved-reasoning
-- spatial-reasoning
 configs:
 - config_name: preview
   data_files:
@@ -19,147 +11,4 @@ configs:
 
 # Instruct-181K
 
-Instruct-181K is the Stage 2 data of PRIMER. It trains
-[ThinkMorph-SSL-Instruct](https://huggingface.co/ThinkMorph/ThinkMorph-SSL-Instruct), starting from
-[ThinkMorph-SSL-Base](https://huggingface.co/ThinkMorph/ThinkMorph-SSL-Base). Each trace has a question, input
-images, textual reasoning with intermediate images, and an answer.
-
-- **Images:** the intermediate images are targets built from the original images: localized
-  regions, depth, geometry, reconstructions, paired views and calibrated objects.
-- **Text:** teacher models write the text; generation and filtering depend on the task.
-
-**Counts.**
-- The paper counts **181,291** traces, the rows of the 21 training folders. This release has
-  **180,490** of them: 142,333 with images and
-  38,157 text only.
-- The difference of 801 rows never reached the model in training:
-  - 800 val rows (`depth_10k`, `egoexo_5k`) have no reasoning image, and Stage 2 drops such
-    rows;
-  - 1 row references a zero-byte image, and the loader skipped it.
-- Stage 2 also trained on the val files, so their rows are part of `train.jsonl`
-  (`source_split` = `val`).
-
-## Composition
-
-| Subset | Rows with images | Text-only rows | Sources |
-| --- | ---: | ---: | --- |
-| `spatial/behind_me_1k` | 1,324 | — | MindCube |
-| `spatial/behind_variants_286` | 284 | — | MindCube |
-| `spatial/directional_query_2k` | 2,594 | — | MindCube |
-| `spatial/movement_direction_1k` | 1,250 | — | MindCube |
-| `spatial/positioned_direction_423` | 423 | — | MindCube |
-| `spatial/turn_degrees_204` | 204 | — | MindCube |
-| `spatial/turn_move_closer_2k` | 2,576 | — | MindCube |
-| `spatial/multiview_58k` | 24,247 | 33,631 | Matterport3D (text only), Realsee3D, 360+x |
-| `spatial/depth_10k` | 9,575 | — | RefSpatial |
-| `spatial/counting_5k` | 5,569 | — | LVIS v1 val (COCO images) |
-| `spatial/rotation_10k` | 10,811 | — | MVImgNet |
-| `spatial/perspective_6k` | 6,512 | — | SAT |
-| `spatial/directional_1k` | 1,162 | — | COCO images (sam3d engine) |
-| `spatial/undirectional_5k` | 4,639 | — | COCO images (sam3d engine) |
-| `spatial/directional_cam_8k` | 8,063 | — | COCO images (Orient-Anything) |
-| `spatial/jigsaw_20k` | 20,000 | — | Pretext-Spatial-50K pools |
-| `perception/jigsaw_20k` | 19,935 | — | jigsaw_20k_v1 pools |
-| `perception/visual_search_bbox_15k` | 14,686 | — | Visual-CoT and crop pools |
-| `perception/visual_search_crop_7.5k` | 7,479 | — | crop pools |
-| `perception/count_1k` | 1,000 | — | synthetic counting scenes |
-| `spatial/egoexo_5k` | 0 | 4,526 | Ego-Exo4D (text only) |
-| **Total** | **142,333** | **38,157** | |
-
-| Functional group (paper) | Pipelines | Paper traces |
-| --- | --- | ---: |
-| L1: visible evidence | visual search, jigsaw QA | 62,100 |
-| L2: spatial structure | counting, depth reasoning | 16,544 |
-| L3: scene reconstruction | panorama, bird's-eye view, alternate view, object rotation, ego-exo transfer | 102,647 |
-| **Total** | | **181,291** |
-
-## Files
-
-| Path | What it is |
-| --- | --- |
-| `shards/SHARDS.json`, `shards/data-*.tar` | The release with images: 142,333 train rows and 286,889 images (157.0 GB), in 148 deterministic tar shards |
-| `text_only/` | 38,157 rows whose images come from Matterport3D or Ego-Exo4D. Their images are not distributed here; see [Rows without images](#rows-without-images) |
-| `preview/*.parquet` | The Dataset Viewer preview: a few rows per subset, with images downscaled to at most 768 px |
-
-## Download and restore
-
-```bash
-python scripts/download_release.py shards/SHARDS.json \
-  --base-url https://huggingface.co/datasets/ThinkMorph/Instruct-181K/resolve/<COMMIT>/shards --output shards
-python scripts/shard_data_release.py restore shards/SHARDS.json shards DATA_ROOT/primer-instruct-181k
-```
-
-(`hf download ThinkMorph/Instruct-181K --repo-type dataset --include "shards/*"` works too.)
-
-- `download_release.py` checks each shard's size and SHA256.
-- `restore` checks every shard before extracting it, then checks every file against the
-  `MANIFEST.json` inside the shards.
-- A restored release holds `train.jsonl` (and `val.jsonl` if present), `images/` and
-  `MANIFEST.json`, which is the layout the BAGEL loader reads.
-- The scripts are in the PRIMER code release.
-
-**Check.** Before upload, the restored release went through the BAGEL loader, which decoded
-every image: 142,333 usable rows and 286,889 images, result
-PASS.
-
-## Row format
-
-```json
-{"id": "<sha256>", "question": "...", "options": "...", "images": ["images/<sha256>.jpg"],
- "reasoning_thought_0": "...", "reasoning_image_0": "images/<sha256>.jpg",
- "reasoning_thought_1": "...", "answer": "...", "subset": "<source subset>", "source_split": "train"}
-```
-
-- The supervision text (`question`, `options`, `answer`/`gt_answer`, every
-  `reasoning_thought_N`) and the image bytes are exactly as in the training copy.
-- Image paths are content-hashed relative paths, and `id` is a public id.
-- `subset` names the storage subset a row came from, and `source_split` the split it had
-  there.
-- Source metadata fields are removed, because they contain internal storage paths.
-
-## Rows without images
-
-`text_only/` holds the rows whose images come from **Matterport3D or Ego-Exo4D**. Their terms do not allow
-us to redistribute the images. Obtain the source data yourself and rebuild the images.
-
-- `text_only/train.jsonl` (and `val.jsonl`): the rows, normalized like the main release
-  (same public ids), with `subset` and `source_split`. The image fields keep the
-  **original file names**.
-- `text_only/images.jsonl`: one record per original image (`subset`, `split`, `name`). Each
-  record gives the byte size and MD5 of the exact file used in training, the upstream
-  dataset and ids, the role and a derivation sentence.
-- `text_only/provenance.jsonl`: per row, the reconstruction parameters (for example yaw
-  angles of paired views, jigsaw grid and tile permutation, original image id and size, take
-  name and frame rank).
-
-To get the source data:
-
-- **Matterport3D:** sign the Matterport3D Terms of Use. Get the form from
-  <https://niessner.github.io/Matterport/> and send it to matterport3d@googlegroups.com. The
-  rows name the Matterport3D panoramas they use.
-- **Ego-Exo4D:** obtain the dataset under its license (<https://ego-exo4d-data.org/>). The
-  frames can be regenerated with `data/builders/egoexo/export_take_ninepack.py` from the PRIMER
-  code release; its README gives the exact parameters. An image name gives the take, ego or
-  exo, and the frame rank.
-
-A rebuilt image may differ in bytes even when it shows the same content. Compare it with the
-recorded MD5.
-
-## Source terms
-
-Each source keeps its own terms. The image release has no blanket license. Matterport3D and
-Ego-Exo4D rows are text only.
-
-## Evaluation proximity
-
-- VStar, MMSI and MindCube are close to the visual-search, panorama and viewpoint
-  supervision.
-- HRBench and MME-RW-L are nearby perception tasks.
-- No evaluation questions or official templates were used to build the data.
-
-## Provenance
-
-- **Training copy.** Training read an internal storage copy that is no longer reachable.
-- **This release.** It is built from the team's archived copy of the same training folders.
-- **What matches.** Row counts agree with the recorded counts of the copy training read.
-- **What could not be checked.** Byte identity with the training copy.
+More details will be added later.