Minimal card for now

README.md

@@ -2,13 +2,6 @@
 license: other
 license_name: per-source-terms
 pretty_name: Pretext-270K
-task_categories:
-- any-to-any
-tags:
-- thinkmorph-ssl
-- primer
-- interleaved-reasoning
-- spatial-reasoning
 configs:
 - config_name: preview
   data_files:
@@ -18,144 +11,4 @@ configs:
 
 # Pretext-270K
 
-Pretext-270K is the Stage 1 data of PRIMER. It trains the Stage 1 model
-[ThinkMorph-SSL-Base](https://huggingface.co/ThinkMorph/ThinkMorph-SSL-Base), from which
-[ThinkMorph-SSL-Instruct](https://huggingface.co/ThinkMorph/ThinkMorph-SSL-Instruct) is trained further on
-Instruct-181K. It has **270,896 instruction-free text-image-text traces**. Each trace starts
-from a corrupted input image; a textual description follows, then the clean target image,
-then a further description.
-
-- **Images:** the target images come from the original source images.
-- **Text:** a teacher model (GPT-5) writes the text segments. It first captions the clean
-  image, then sees the caption and both images, so it knows the target, which the student
-  does not see in its input.
-
-Stage 1 trains on the train rows only: 270,896 rows here, which is the paper's count. The
-1,600 val rows were not used for training.
-
-## Composition
-
-| Subset | Rows with images | Text-only rows | Sources |
-| --- | ---: | ---: | --- |
-| `mixed_100k` | 100,000 | — | VSI-590K pools (YouTube room tours, Hypersim, ARKitScenes, ScanNet, ...), LLaVA-OneVision-1.5 pools, DeepVision-103K |
-| `jigsaw_fragments_18k` | 8,085 | 10,266 | Matterport3D (text only), Realsee3D, 360+x |
-| `sa1b_55k` | 55,442 | — | SA-1B |
-| `general_99k` | 98,703 | — | PixMo-Cap |
-| **Total** | **262,230** | **10,266** | |
-
-### Counts by transformation family (train rows)
-
-| Family | Rows | Paper |
-| --- | ---: | ---: |
-| Jigsaw restoration | 52,975 | 52,975 |
-| Colorization | 36,104 | 36,104 |
-| Super-resolution | 36,409 | 36,409 |
-| Masked inpainting | 35,950 | 35,950 |
-| Rotation/flip recovery | 36,763 | 36,763 |
-| Denoising | 36,023 | 36,023 |
-| Deformation recovery | 36,672 | 36,672 |
-| **Total** | **270,896** | **270,896** |
-
-### Counts by source (train rows) and the 443-row difference
-
-The paper's source table adds up to **270,453**, which is **443 fewer** than the family total
-of 270,896. The data explains the gap:
-
-| Source | Rows in data | Paper |
-| --- | ---: | ---: |
-| PixMo-Cap | 98,303 | 98,303 |
-| SA-1B | 55,042 | 55,042 |
-| `mixed_100k`: VSI-590K + LLaVA-OneVision-1.5 + DeepVision-103K | 99,600 | 58,546 + 27,379 + 13,675 = 99,600 |
-| Matterport3D | 10,035 | 10,035 |
-| Realsee3D | 7,251 | 7,251 |
-| 360+x (`source` = `360x_pairs_merged`) | 222 | 222 |
-| 360+x (`source` = `stage0`) | 443 | not listed |
-| **Total** | **270,896** | **270,453** |
-
-- **The 443 rows.** They are 360+x images whose `source` field reads `stage0`.
-- **DeepVision.** It can be counted directly in `mixed_100k`: 8,455 + 5,220 = 13,675.
-
-## Files
-
-| Path | What it is |
-| --- | --- |
-| `shards/SHARDS.json`, `shards/data-*.tar` | The release with images: 260,861 train rows, 1,369 val rows and 533,366 images (140.9 GB), in 133 deterministic tar shards |
-| `text_only/` | 10,266 rows whose images come from Matterport3D. 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/Pretext-270K/resolve/<COMMIT>/shards --output shards
-python scripts/shard_data_release.py restore shards/SHARDS.json shards DATA_ROOT/primer-pretext-270k
-```
-
-(`hf download ThinkMorph/Pretext-270K --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: 262,230 usable rows and 533,366 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**. 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.
-
-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.
-
-- **Code license.** SAM's Apache-2.0 code license does not cover the SA-1B images.
-- **360+x.** Its project page shows CC BY-NC-SA 4.0.
-- **Realsee3D.** Its Zenodo record lists CC BY 4.0.
-- **Matterport3D.** Its rows are text only.
-
-## 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.** Train row counts, family counts and source counts match the paper's
-  records.
-- **What could not be checked.** Byte identity with the training copy.
+More details will be added later.