# Training sets of the offline-RL experiments

These are the six GRPO training sets behind the paper's training-data table. They are ready to train on. Each
set is 2,400 policy prompts, i.e. 16 prompts per step for 150 steps (one epoch) or 300 steps (two epochs).

| Set | UI-Genie | t | t-1 | Rows |
|---|---:|---:|---:|---:|
| `ui100` | 2,400 | – | – | 2,400 |
| `t100` | – | 2,400 | – | 2,400 |
| `ui75_t25` | 1,800 | 600 | – | 2,400 |
| `ui50_t25_t1_25` | 1,200 | 600 | 600 | 2,400 |
| `ui50_t35_t1_15` | 1,200 | 840 | 360 | 2,400 |
| `ui40_t35_t1_25` | 960 | 840 | 600 | 2,400 |

`t100` has 2,080 unique contexts plus 320 repeated ones. A repeated row carries a `__repN` suffix on its
`sample_id`.

## Files

```
training_sets/
  <set>/ms_swift_grpo.jsonl   one row per prompt, in training order
  images/uigenie/             8,534 screenshots (UI-Genie contexts)
  images/af_t/                7,262 screenshots (AndroidFlux t contexts)
  images/af_t_minus_1/        2,181 screenshots (AndroidFlux t-1 contexts)
  sha256.json                 row count and SHA-256 of each ms_swift_grpo.jsonl
  sources/                    label-bearing t / t-1 samples the sets were built from (see Provenance)
```

The six sets share their screenshots: 17,977 files, 8.2 GB in total. Each file is named by its SHA-256.

Each row has these parts:

- `messages` and `images`: the policy prompt. This is the Qwen3-VL agent prompt with the latest 10 history steps
  and the current screenshot.
- `reward_context`: the reward-model input. It holds the UI-Genie layout, the last ≤5 screens, the
  screen resolution, the reference action and, for AndroidFlux rows, the `is_next_terminate` label.

Screenshot paths in both parts are relative to `training_sets/`, e.g.
`images/af_t/<sha256>.png`.

## Training on a set

Rewrite the relative screenshot paths to absolute ones, then point the trainer at the file:

```bash
cd training_sets
sed "s#\"images/#\"$PWD/images/#g" ui40_t35_t1_25/ms_swift_grpo.jsonl > ui40_t35_t1_25.abs.jsonl
```

The training code and configs are in the AndroidFlux repository under `rl/train/`:

- `rl/train/README.md`: one-command training and evaluation.
- `code/replay_recovery_benchmark/rl_training/configs/`:
  - `b3_*` and `b4_*`: 150 steps.
  - `b5_*_e2`: 300 steps.

Pass the file as `DATASET` (or `training.dataset`).

## Provenance

The sets are built from these sources:

- `Gyubeum/UI_Genie_128k` @ `d7987e1f`.
- The `t` / `t_minus_1` slices of this repository: their `context.parquet` and `images.parquet`, and
  `androidflux_tools/export_repaired_interchange.py`.
- `sources/t/samples.parquet` and `sources/t_minus_1/samples.parquet`: the `t` / `t_minus_1` samples with the
  `is_next_terminate` and goal-match labels as they were when the sets were built. The top-level slices carry
  later label corrections, which differ in 11 `is_next_terminate` values.

AndroidFlux contexts whose goal does not match their source trajectory are dropped. No rollout or reward-model
filtering is applied.

`rl/train/data_curation/reproduce.sh` rebuilds all six sets from those sources. These files are the ones the
runs trained on. Their content matches that rebuild hash for hash; only the screenshot root is rewritten to the
relative form above.

`curated_set/` in this repository lists the same rows as indices into the published slices.