---
task_categories:
- visual-question-answering
- image-to-text
tags:
- chain-of-thought
- depth-estimation
- edge-detection
- segmentation
- webdataset
size_categories:
- 500K<n<1M
---

# CoVT_scaleup

Chain-of-thought VQA data for training a model to generate auxiliary depth /
edge / segmentation maps as intermediate reasoning steps inside a `<think>`
block, e.g.:

```
<think> The depth map of the image is [depth map], The edge map of the image
is [edge map], ... The segmentation of the image is [seg map]. </think>
The answer is X
```

Companion code: [Claimentine/CoVT_scaleup on
GitHub](https://github.com/Claimentine/CoVT_scaleup) (BAGEL fine-tuning
pipeline, ThinkMorph-style boundary signaling — see that repo's README for
the modeling side).

## Stats

- 593,790 records
- 297 WebDataset shards (`shards/shard-00000.tar` … `shards/shard-00296.tar`)
- 1,942,181 images total (original RGB photo + generated depth map + PiDiNet
  edge map +, for a subset of records, a SAM3 segmentation map)
- ~332 GB total

## Format

Each shard is a plain tar archive following the [WebDataset](https://github.com/webdataset/webdataset)
convention: files sharing the same zero-padded 9-digit key belong to one
sample.

```
000000000.json          # metadata + conversation for this sample
000000000.img0.png      # original RGB photo
000000000.img1.png      # depth map
000000000.img2.png      # PiDiNet edge map
000000000.img3.png      # SAM3 segmentation map (present for most, not all, records)
000000001.json
000000001.img0.png
...
```

`{key}.json` is the original CoT record with `image` rewritten to point at
the in-tar filenames instead of absolute cluster paths:

```json
{
  "id": "identity_177477",
  "image": ["000000000.img0.png", "000000000.img1.png", "000000000.img2.png", "000000000.img3.png"],
  "conversations": [
    {"from": "human", "value": "<image>\nQuestion: ..."},
    {"from": "gpt", "value": "<think> The depth map of the image is <image>, The edge map of the image is <image>, ... </think> The answer is ..."}
  ],
  "seg_coverage": 0.3558
}
```

`image[0]` is always the original RGB photo; `image[1:]` are the auxiliary
maps in the same order they're referenced by the `<image>` tags inside the
`<think>` block of the `gpt` turn. Some records only have 3 images (no
segmentation map) — the number of `<image>` tags in the `gpt` turn always
matches `len(image) - 1`.

Each shard also has a `shard-XXXXX.manifest.json` (record count / image
count for that shard — packing-time bookkeeping, not needed to load the
data).

## Loading

```python
import webdataset as wds

url = "https://huggingface.co/datasets/claimentine/CoVT_scaleup/resolve/main/shards/shard-{00000..00296}.tar"
ds = wds.WebDataset(url).decode("pil")

for sample in ds:
    record = sample["json"]
    rgb = sample["img0.png"]
    depth = sample["img1.png"]
    ...
```

Or download shards directly and use the loader in the companion code repo
(`data/cot_dataset.py:CoTJSONLIterableDataset` expects a plain jsonl +
loose image files on disk — extract shards locally first if you want to
train with that exact loader unmodified).

## Provenance / license note

Source questions/images are drawn from a mixture of existing VQA datasets
(TallyQA, CLEVR, A-OKVQA, ST-VQA, and others reachable via the record `id`
prefix, e.g. `cauldron/tallyqa/...`), each carrying their own upstream
license — this repo does not re-license them. Depth maps, PiDiNet edge
maps, and SAM3 segmentation maps are generated/derived by this project. No
explicit license is declared at the dataset-repo level; treat as
research-use data and check the upstream source dataset's license before
any other use.