README.md

---
license: cc-by-nc-2.0
configs:
- config_name: samples
  data_files:
  - split: train
    path: samples/*.parquet
- config_name: images
  data_files:
  - split: train
    path: images/*.parquet
- config_name: synthetic_attribute_samples
  data_files:
  - split: first_1000
    path: synthetic_attribute/samples/first_1000.parquet
  - split: remaining_4000
    path: synthetic_attribute/samples/remaining_4000.parquet
- config_name: synthetic_attribute_images
  data_files:
  - split: first_1000
    path: synthetic_attribute/images/first_1000/*.parquet
  - split: remaining_4000
    path: synthetic_attribute/images/remaining_4000/*.parquet
---

# Visual_Agent Parquet

Parquet distribution of `albert13200/Visual_Agent` with 3679 tool-use
trajectories and 4816 deduplicated images.

## Tables

- `samples`: one row per trajectory. `record_json` preserves the complete original row.
- `images`: one row per unique relative image path, with binary bytes, MIME type,
  dimensions, file size, and SHA-256.
- `synthetic_attribute_samples`: 5,000 synthetic SAM3 crop-replay attribute
  trajectories, kept separate as `first_1000` and `remaining_4000` splits.
- `synthetic_attribute_images`: the 10,000 referenced synthetic input/crop images,
  partitioned to match the two synthetic sample splits.

Separating the tables prevents the same image bytes from being embedded repeatedly when
multiple trajectories reference one image.

## Use directly

```python
from datasets import load_dataset

samples = load_dataset("albert13200/Visual_Agent_Parquet", "samples", split="train")
images = load_dataset("albert13200/Visual_Agent_Parquet", "images", split="train")

synthetic_first = load_dataset(
    "albert13200/Visual_Agent_Parquet",
    "synthetic_attribute_samples",
    split="first_1000",
)
synthetic_remaining = load_dataset(
    "albert13200/Visual_Agent_Parquet",
    "synthetic_attribute_samples",
    split="remaining_4000",
)
```

## Restore the path-based training layout

```bash
uvx --from huggingface-hub hf download \
  albert13200/Visual_Agent_Parquet \
  --repo-type dataset \
  --local-dir Visual_Agent_Parquet

uv run --with pyarrow \
  Visual_Agent_Parquet/materialize_visual_agent_parquet.py \
  --input-root Visual_Agent_Parquet \
  --output-root Visual_Agent
```

This writes `Visual_Agent/training_trajectories_natural/` with the combined JSONL and
all relative image files expected by the existing training pipeline. Image bytes are
validated against their stored SHA-256 before use.

## Restore the synthetic attribute set

```bash
uv run --with pyarrow \
  Visual_Agent_Parquet/synthetic_attribute/materialize_synthetic_attribute_parquet.py \
  --input-root Visual_Agent_Parquet/synthetic_attribute \
  --output-root Synthetic_Attribute
```

Use `--partition first_1000` or `--partition remaining_4000` to restore only one
synthetic partition. Without `--partition`, the script restores all 5,000 rows and
verifies that the combined JSONL exactly matches the source SHA-256.