---
dataset_info:
  features:
  - name: messages
    list:
    - name: content
      dtype: string
    - name: role
      dtype: string
  - name: ground_truth
    dtype: string
  - name: dataset
    dtype: string
  - name: env_config
    struct:
    - name: env_name
      dtype: string
    - name: image
      dtype: string
    - name: task_id
      dtype: string
  - name: source
    dtype: string
  splits:
  - name: train
    num_bytes: 133363912
    num_examples: 37484
  download_size: 29005927
  dataset_size: 133363912
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
license: cc-by-4.0
language:
- en
tags:
- tmax
- terminal
- agents
- reinforcement-learning
---

# Recursive-Task-Synthesis for tmax

**Images require building:** the complete dataset and build contexts are included. Image builds are deferred; run the resumable script below before using these environments.

All **37,484** task directories from [Zhongzhi1228/Recursive-Task-Synthesis](https://huggingface.co/datasets/Zhongzhi1228/Recursive-Task-Synthesis), pinned to `be44f96808d5a9b599d5cb024341ff00091adeb7`, converted to tmax's `swerl_vanillux_sandbox` format.

The train split uses the same `messages`, `ground_truth`, `dataset`, `env_config`, and `source` schema as the other `hamishivi/agent-task-*` datasets. Messages use the tmax Vanillux templates; `dataset` is `passthrough`. Task IDs are prefixed with `recursive_task_synthesis__` to avoid collisions when combining task archives. `task-manifest.json` maps them back to original paths.

`task-data.tar.gz` contains a directory per task with all original task files preserved byte-for-byte, plus `image.txt`. Instructions, verifiers, build contexts, metadata, and reference solutions are retained. The tmax loader exposes only environment seed files and defers tests until submission; reference solutions are not mounted into the agent sandbox.

## Usage

```python
from datasets import load_dataset
from huggingface_hub import hf_hub_download

train = load_dataset("hamishivi/agent-task-recursive-task-synthesis", split="train")
archive = hf_hub_download("hamishivi/agent-task-recursive-task-synthesis", "task-data.tar.gz", repo_type="dataset")
```

Extract the archive and set the sandbox `task_data_dir` to the directory containing the task directories. Every row specifies its image explicitly. Task resource requirements and verifier timeouts remain in the original `task.toml`; configure the training harness accordingly.

## Images

Images use `hamishi740/agent-task-recursive-task-synthesis:<source-environment-hash>` and target Linux AMD64. See `task-manifest.json` for build status.

## Validation and attribution

All rows passed schema and Parquet round-trip checks, and every archived task file was checked against the downloaded source bytes. See `validation.json`. This does not constitute an evaluation of every task or a full training run.

The upstream license is cc-by-4.0. Consult the upstream dataset card for citation and provenance. tmax Vanillux prompts are adapted from mini-swe-agent (MIT).

## Reproducibility and runtime coverage

This conversion targets `hamishivi/tmax-private`, branch `geomean_mask`, commit `9fff7af6e86f2589830637772ff5b10b1c080131`. All original task files are preserved; added task IDs, training rows, and image references adapt the packaging for tmax. See `UPSTREAM_README.md` for the source authors' attribution and citation details.

Validation covers all task records and archived file contents. No container build or runtime task evaluation was performed for this release. Upstream validation claims do not replace testing in your training harness.

## Resumable image builds

Requires Docker Buildx and a Docker Hub login with write access to the target image repository. The script skips published tags and digest-pinned upstream images, saves per-image logs, and retries unfinished builds on the next run.

```sh
mkdir -p tasks
tar -xzf task-data.tar.gz -C tasks
python build_dataset_images.py --task-data-dir tasks --manifest task-manifest.json --workers 8 --timeout 1800
```

Images built by this script target Linux AMD64. If using another Docker Hub namespace, update image references consistently in the manifest, training records, and task `image.txt` files. The image status file is a publication-time snapshot.

The source includes 67 references to three unavailable local image names. These tasks use the supplied build contexts and new image targets; the original names are retained in the manifest as `unavailable_upstream_image`. Source task-group IDs are also retained for grouping and splitting related tasks.

## Apptainer images

The current Apptainer pool and unified download manifest are maintained in [TMaxxx/agent-task-recursive-task-synthesis](https://huggingface.co/datasets/TMaxxx/agent-task-recursive-task-synthesis). New SIF uploads go to TMaxxx; earlier images remain available here. Use the downloader and manifest in the linked repository to retrieve all available images with tmax-compatible filenames.