---
license: other
license_name: appgen-generated-data-notice
license_link: LICENSE.md
task_categories:
  - reinforcement-learning
tags:
  - gui-agent
  - mobile-agent
  - android
  - reinforcement-learning
  - ued
  - grpo
  - reproducibility
pretty_name: AppGen UED-GRPO v32r12 Training Data and State
---

# AppGen UED-GRPO v32r12 training bundle

This dataset is the reproducibility bundle for
`uedgpo_q3_ngc_amex_pbrs_v32r12`. It fills the release gap left by the
previous static-data and model repositories: the run's online-generated
environments and its curriculum teacher states.

The bundle is designed to be materialized together with the already published,
revision-pinned static inputs. No public benchmark, AndroidWorld evaluation,
or held-out evaluation screenshots are included.

## Contents

- `artifacts/v32r12_online_envs.tar.gz`: all 19 online-generated environments
  produced by the run, including `ui_structure.json`, HTML, rendered PNGs, and
  generation metadata.
- `artifacts/v32r12_static_envs.tar.gz`: all 150 static UED environments used by
  the run, packed for fast one-file transfer.
- `teacher_states/global_step_*/teacher_state.json`: curriculum/replay state at
  every saved checkpoint from step 10 through step 110.
- `run/`: exact curriculum seed, launcher registry, policy system prompt, and
  Kubernetes run spec.
- `manifests/`: per-file SHA-256 hashes, per-environment inventory, static-input
  hashes, dependency revisions, and the release manifest.
- `download_and_materialize.py`: one-command materializer and verifier.

The v32r12 static pool contains 150 environments under
`preload_v5/{base,easy,uniform}`. Twelve are pinned warm-up seeds and 138 remain
in the runtime unseen pool after seed exclusion. The same byte-identical files
remain in `luca0621/appgen-training-data` at the pinned provenance revision, but
are also bundled here to avoid enumerating thousands of small files from that
large repository during materialization.

## Quick start

Install the downloader, fetch this small script, and materialize the complete
training-data layout:

```bash
pip install -U "huggingface_hub>=0.28.0"
hf download luca0621/appgen-ued-v32r12-training-data \
  download_and_materialize.py --repo-type dataset --local-dir v32r12-release
python v32r12-release/download_and_materialize.py \
  --output-dir ./appgen-v32r12-data
```

The default command downloads and verifies:

1. this release's online environments and teacher states;
2. the bundled 150 static environments, verified against the pinned
   `appgen-training-data` revision;
3. the exact `sft_qwen3_UNIFIED.json` used by the UED environment loader; and
4. the launcher registry and run metadata.

To also download the 18-GB initializer model:

```bash
python v32r12-release/download_and_materialize.py \
  --output-dir ./appgen-v32r12-data --include-model
```

After completion, the output reproduces the original relative layout beneath
`/data/appgen`, including:

```text
appgen-v32r12-data/
├── preload_v5/{base,easy,uniform}/
├── sft_qwen3_UNIFIED.json
├── training_env_pool/_launcher_registry.yaml
├── verl_q3_ngc_amex_pbrs_v32r12_pool/
│   ├── envs/
│   ├── grpo_curriculum_seed_category10_v28.json
│   └── system_prompt.sft_exact.txt
└── verl_q3_ngc_amex_pbrs_v32r12_ckpts/
    └── global_step_*/teacher_state.json
```

## Pinned dependencies

| Role | Repository | Revision |
|---|---|---|
| Static UED environments | `luca0621/appgen-training-data` | `fed731b4dfa58118cb3014cf2656a42b6ae92f0d` |
| SFT goal manifest | `luca0621/appgen-sft-data` | `c195ae15abd3d6aaa07f971b8d732e7a28fa8dbf` |
| NGC SFT source | `luca0621/appgen-sft-ngc-v1` | `769ea99dbc4ff190048ae0db37eb6310dba595e0` |
| AMEX materializer input | `Yuxiang007/AMEX` | `17196b29c88dd48a7fb90ef9131bc5c7bf39f26e` |
| UED initializer model | `namhokaist/appgen-qwen3-vl-8b-sft-ngc-amex-avariant-E-ngc-lr2p5e7-1ep` | `6cdf0aa413850771f9a6f4c4da38f53d9d060f1c` |

The initializer's audited materialization receipt reports
`ngc_primary_rows=2768` and `amex_rows=0` for arm E. AMEX is therefore a pinned
materializer/provenance input but contributes no selected gradient row to this
specific initializer. The selected NGC configuration has 3,588 exposures and
3,168 unique semantic examples.

## Online-environment usage

All 19 generated environments are released, including those that were only
generated and never retained in a saved replay buffer. At step 100, three
online environments appear in the teacher buffer and each has one visit, which
is score-only under the run contract. At step 110, one online environment has
two visits and was therefore replayed for a gradient update. Raw teacher states
are included so downstream users can audit this distinction directly.

## Integrity

`download_and_materialize.py` verifies the archive SHA-256, every extracted
online-environment file, all 150 static `ui_structure.json` files, the exact SFT
JSON, every teacher state, and the run sidecars. The top-level release manifest
records the same digests and source revisions.

## Models

The corresponding public checkpoints include:

- `luca0621/appgen-qwen3-uedgrpo-ngc-amex-pbrs-v32r12-step100`
  at revision `58be2eec39b1f31291347b9153e55d7ccec91b5d`
- `luca0621/appgen-qwen3-uedgrpo-ngc-amex-pbrs-v32r12-step110`
  at revision `72f05813148c2f1174447490b15abb3a9abb0445`

Model repositories contain weights and model provenance. This dataset contains
the environment and curriculum state needed to reproduce and audit training.