README.md

---
language:
- en
task_categories:
- text-generation
tags:
- agent
- code
- agentic-tasks
- harbor
- reinforcement-learning
configs:
- config_name: default
  data_files:
  - split: train
    path: "*/tasks.parquet"
- config_name: archive
  data_files:
  - split: train
    path: "*/archive.parquet"
---

# Validated tasks

Agent tasks in [Harbor](https://github.com/marin-community/harbor) format, together with the result of validating them. The validation is described in the [OT-next-data validation README](https://github.com/MLI-lab/OT-next-data/blob/main/validation/README.md).

## Layout

Each data source has its own folder with three files:

| File | Content |
| --- | --- |
| `README.md` | what the tasks are and where they come from |
| `tasks.parquet` | the tasks that are kept |
| `archive.parquet` | the tasks that were excluded, with the reason |

`runs/` holds one file per validation run: the contract of the run and the number of tasks kept and archived at each stage.

Load the kept tasks, or the archived ones:

```python
from datasets import load_dataset
kept = load_dataset("FWeindel/validated-tasks", split="train")
archived = load_dataset("FWeindel/validated-tasks", "archive", split="train")
```

## Download one datasource and its images

To run one datasource, download its original Parquet. Replace `SOURCE` with
its folder name and choose a local workspace:

```python
from huggingface_hub import hf_hub_download

parquet = hf_hub_download(
    repo_id="FWeindel/validated-tasks",
    repo_type="dataset",
    filename="SOURCE/tasks.parquet",
    local_dir="/path/to/workspace/dataset",
)
```

This downloads the tasks only. When you run that file with the
[OT-next-data launcher](https://github.com/MLI-lab/OT-next-data/blob/main/validation/README.md),
its [image downloader](https://github.com/MLI-lab/OT-next-data/blob/main/validation/publishing/image_release.py)
automatically downloads the published Apptainer images needed by your selected
tasks, verifies them, and caches them for later runs. You do not need to download
other datasources' images.

This works for releases with image references in their Parquet metadata.
Keep the original file to preserve those references. Tasks without published
images are built as usual. See the
[download and cache guide](https://github.com/MLI-lab/OT-next-data/blob/main/validation/docs/image-publication.md)
for details and Docker support.

## Columns

| Column | Meaning |
| --- | --- |
| `path` | task ID |
| `task_binary` | the task as a tar archive, as validated |
| `content_sha256` | hash of the task's files |
| `stages_passed` | the validation stages the task has passed, e.g. `1,3,4,5` |
| `archive_stage` | the stage that excluded the task; empty for kept tasks |
| `archive_reason` | why the task was excluded; empty for kept tasks |
| `run` | the file in `runs/` of the run that decided this |

A changed task has a new `content_sha256` and is validated again from stage 1.

## Keep or archive

| Stage | Rule |
| --- | --- |
| 1, static checks | archive if any check fails |
| 3, build | archive if the container does not build or start |
| 4 and 5, oracle and NOP | archive if the reward is wrong: the reference solution must get 1, doing nothing must get 0 |
| 2, 6, 7 and 8 | never archive; a person judges |

## Updates

Each validation run is proposed as one pull request. It adds the run file, updates the two Parquets of each affected data source, and states in its description how many tasks were kept and archived, and why.