---
license: other
license_name: mixed-see-per-dataset-license
pretty_name: skilltrainbench training tasks
task_categories:
  - other
tags:
  - agents
  - agentic-benchmark
  - harbor
  - training-split
configs: []
---

# skilltrainbench training tasks

The **training half** of the skilltrainbench benchmark suite: for each of the
four datasets, the `dev_task_names` of its pinned train/test split, in Harbor
task format.

> **The held-out/test tasks are not in this repository.** Neither are the
> published splits that are not the pin, nor the tasks that fall outside each
> pinned split's task set. Use this repository for skill formation and
> training; evaluate on the held-out half, which stays in the private source
> repository.

| Dataset | Pinned split | Task set | Train tasks here | Held-out (absent) |
|---|---|---|---:|---:|
| [`healthbench`](healthbench) | `hb-hardbeh-s002` | `hard_behavioural` | 200 | 160 |
| [`hle`](hle) | `hle-stem517-s001` | `stem517` | 412 | 105 |
| [`QuantitativeFinance-Bench`](QuantitativeFinance-Bench) | `qf-medhard-strat-80-s005` | `medium_hard` | 54 | 14 |
| [`tau3-bench`](tau3-bench) | `tau3-banking-70-s001` | `banking_knowledge` | 67 | 30 |
| **Total** | | | **733** | **309** |

## Layout

```
<dataset>/
  README.md        dataset card: scoring, judge configuration, provenance
  LICENSE          upstream licence, where the upstream declares one
  dataset.toml     manifest, filtered to the tasks present here
  tasks/<name>/    one directory per training task
splits/
  README.md        the pin table, learner pools and tool-call cutoffs
  <dataset>/       the pinned split file, reproduced byte-for-byte
```

Each split file keeps both `dev_task_names` and `heldout_task_names` so that
its `assignment_sha256` still verifies against the digests in
[`splits/README.md`](splits/README.md). The held-out names therefore resolve to
directories that are not in this repository.

`dataset.toml` in each dataset folder has been filtered to the tasks present;
every other file is byte-identical to the source repository.

## Download

Everything:

```bash
hf download armin-aptura/skilltrainbench-public --repo-type dataset --local-dir .
```

One dataset and its split:

```bash
hf download armin-aptura/skilltrainbench-public --repo-type dataset \
  --include "tau3-bench/**" --include "splits/tau3-bench/**" --local-dir .
```

## Licensing

Licensing is per dataset folder, not uniform across this repository. Check the
folder you intend to use.

| Dataset | Licence | Attribution |
|---|---|---|
| `QuantitativeFinance-Bench` | CC BY-NC 4.0 (`LICENSE`) | QF-Bench |
| `healthbench` | MIT (`LICENSE`) | OpenAI |
| `tau3-bench` | MIT (`LICENSE`) | Sierra Research |
| `hle` | no `LICENSE` file in this folder | `harbor-framework/harbor-datasets` |

`QuantitativeFinance-Bench` is non-commercial.

`QuantitativeFinance-Bench` tasks carry a canary GUID, documented in
[its README](QuantitativeFinance-Bench/README.md), so that contamination of a
model's training corpus can be detected.

## Provenance

Derived from the private repository `skilltrainbench/dataset` at revision
`b1d7426f376c6310246323801e8d490582d39276`. Each dataset folder's own README
records the upstream source, generator commit and task-tree digest.