README.md

---
license: apache-2.0
dataset_info:
  features:
  - name: example_id
    dtype: string
  - name: watchface_id
    dtype: string
  - name: watchface_name
    dtype: string
  - name: watch_time
    dtype: string
  - name: time_format
    dtype: string
  - name: shows_seconds
    dtype: bool
  - name: hour_visible
    dtype: int32
  - name: minute_visible
    dtype: int32
  - name: second_visible
    dtype: int32
  - name: text
    dtype: string
  splits:
  - name: train
    num_bytes: 130732460
    num_examples: 800269
  - name: validation
    num_bytes: 4180551
    num_examples: 25600
  download_size: 69712029
  dataset_size: 134913011
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
---

# PixMo-Clocks
PixMo-Clocks is a collection of virtual watch faces and time annotations. 
The repo supplies the meta-data needed to build the data but does not directly contain the images, which are from facer.io.

PixMo-Clocks is a part of the PixMo dataset collection and was used to train the [Molmo family of models](https://huggingface.co/collections/allenai/molmo-66f379e6fe3b8ef090a8ca19)

Quick links:
- 📃 [Paper](https://molmo.allenai.org/paper.pdf)
- 🎥 [Blog with Videos](https://molmo.allenai.org/blog)


## Loading

```python
clocks_dataset = datasets.load_dataset("allenai/pixmo-clocks")
```

## Data Format
The data includes:
- `watchface_id`: The watchface id to use to generate the image
- `watchface_name`: The name of the watchface used
- `watch_time`: The time the watch was set to, not all the details of the time will be visible on the watchface, so the
the following two fields are needed.
- `time_format`: The format of the watch, can be:
  - `"No time visible`": The time cannot be read at all, we still include these examples as no-answer training examples
  - `"Without AM/PM`": AM/PM cannot be determined
  - `"With AM/PM`": AM/PM can be determined (either because the watch shows military time, or shows a AM/PM indicator somehow)
- `shows_seconds`: Whether seconds are shown

For convenience, this dataset includes a few fields derived from this data about what is visible on the watch:
- `hour_visible`: The hour visible on the watch, between 0 and 23, -1 means not visible, 0 is 12:00am.
  If AM/PM cannot be determined the hour will be between 0 and 11
- `minute_visible`: The minute the watch should be set to, between 0 and 59, -1 means not visible
- `second_visible`: The second the watch should be set to, between -1 and 59, -1 means not visible
- `text`: A text string of the time

## Generating Images
Generating the image requires setting the watch face to the correct time and then downloading the image from facer.io.
We will include a script to generate the images as part of our code release.

## Splits
The data is divided into validation and train splits. These splits are ``unofficial`` because we do not use this data for evaluation anyway. However,
they reflect what was used when training the Molmo models, which were only trained on the train split.

## License
This dataset is licensed under Apache 2.0. It is intended for research and educational use in accordance with Ai2's [Responsible Use Guidelines](https://allenai.org/responsible-use).