README.md

# eval3_objectvla_vl_pairs

ObjectVLA-style vision-language co-training data for the Eval 3 face/celebrity
identification task. Companion dataset to
[`HBOrtiz/so101_eval3_aug_v3_200celebs`](https://huggingface.co/datasets/HBOrtiz/so101_eval3_aug_v3_200celebs).

## Contents

- **`manifest.parquet`** — 176 670 VL pairs (one row per (frame, portrait, caption_type)).
- **`data.tar.zst`** — 1.7 GB compressed archive containing:
  - `images/chunk-{000..003}/*.jpg` — 29 445 wrist-cam frames (480 × 640, 3 sampled per episode at 20% / 50% / 80%).
  - `references/<episode>__ref.jpg` — 9 815 reference photos (480 × 480, one per episode — the "what the target should look like" cue).
  Extract with `tar --zstd -xf data.tar.zst` (creates the two subdirs).
- **`_stats.json`** — dataset statistics + skip counters.

## Manifest schema

| column | type | meaning |
|---|---|---|
| `image_path` | str | path to wrist-cam JPEG (e.g. `images/chunk-001/quick_lecun_LSO_ep01_..._var00__f0107.jpg`) |
| `reference_image_path` | str | path to reference JPEG (e.g. `references/quick_lecun_LSO_ep01_..._var00__ref.jpg`) |
| `prompt` | str | the input prompt |
| `target` | str | the expected completion |
| `bbox_xyxy_norm` | list[float] | bounding box as `[x1, y1, x2, y2]` normalized to [0, 1] in (W, H) of the wrist-cam image |
| `celeb_name` | str | celebrity display name (e.g. `"Yann LeCun"`) |
| `celeb_slug` | str | slug form (e.g. `"yann_lecun"`) |
| `caption_type` | str | `"location_explicit"` or `"qa_grounded"` |
| `episode` | str | source episode in the robot dataset |
| `frame_idx` | int | frame index in the source mp4 |
| `pid` | int | portrait id (0, 1, or 2) |

## Two camera streams per sample

Mirroring the original Eval 3 robot dataset's two-camera setup:

- **`image_path`** is the **wrist-cam** view — 480 × 640 of the workspace with 3 visible printed portraits. This is the perception domain the robot acts in.
- **`reference_image_path`** is the **reference** view — 480 × 480 of the target celeb's photo, the same image for all frames of a given episode (it's a constant-frame video in the source robot dataset).

The bbox in `bbox_xyxy_norm` is the **paper-region bbox in the wrist-cam image** (where the target portrait sits on the table), NOT the reference image (which is whole-frame). Use the reference image as auxiliary visual context, not for bbox grounding.

## Two caption types

Per visible portrait per frame we emit one row of each type (so 6 rows per
frame for 3 visible portraits):

**`location_explicit`** — model is asked what's in the image, expected to emit name + bbox:
```
prompt: "What is in this image?"
target: "The printed photo of Barack Obama is at [0.06, 0.43, 0.41, 0.90]."
```

**`qa_grounded`** — model is given a bbox in the prompt, expected to name the person there:
```
prompt: "Who is in the printed photo at [0.06, 0.43, 0.41, 0.90]?"
target: "Barack Obama"
```

## Why this format

Following the ObjectVLA recipe ([arxiv 2502.19250](https://arxiv.org/abs/2502.19250) §3),
the bbox is **whole-object** (the printed portrait rectangle), computed
from the augmentation pipeline's cached `portrait_corners.json` (no new
face-detection inference required). Loose-crop / whole-object boxes
empirically outperform tight face crops for identity discrimination
([Banerjee 2022](https://arxiv.org/abs/2208.02991), LFW ablations).

## Stats

- 192 unique celebrities
- 9 815 source episodes (of 9 842 in the robot dataset — 27 with missing corners metadata)
- 29 445 sampled wrist-cam frames (3 per episode, at 20% / 50% / 80% time fractions)
- 9 815 reference frames (1 per episode — constant-frame video, single sample taken)
- 176 670 VL pairs (50 / 50 split between `location_explicit` and `qa_grounded`)

## Intended use

Mix into Pi0.5 fine-tuning at a 10:1 robot:VL batch ratio per
[ObjectVLA](https://arxiv.org/abs/2502.19250). See
[`eval_3/tracks/TRACK_OBJECTVLA.md`](https://github.com/Ace3Z/LeMonkey/blob/main/eval_3/tracks/TRACK_OBJECTVLA.md)
in the source repo for the full training recipe.

## Loading example

```python
import subprocess
from huggingface_hub import hf_hub_download
import pyarrow.parquet as pq

# Download + extract the image archive
tar = hf_hub_download("HBOrtiz/eval3_objectvla_vl_pairs",
                        "data.tar.zst", repo_type="dataset")
subprocess.run(["tar", "--zstd", "-xf", tar])  # creates images/ and references/

# Load manifest
manifest = pq.read_table(hf_hub_download(
    "HBOrtiz/eval3_objectvla_vl_pairs", "manifest.parquet", repo_type="dataset"
)).to_pandas()

# Sample one row
row = manifest.iloc[0]
print(row.prompt, "→", row.target)
print(f"wrist-cam: {row.image_path}")
print(f"reference: {row.reference_image_path}")
```