| --- |
| configs: |
| - config_name: ui_genie |
| data_files: |
| - split: train |
| path: ui_genie/samples.parquet |
| - config_name: t_minus_n |
| data_files: |
| - split: train |
| path: t_minus_n/samples.parquet |
| - config_name: successful |
| data_files: |
| - split: train |
| path: successful/samples.parquet |
| - config_name: t_minus_1 |
| data_files: |
| - split: train |
| path: t_minus_1/samples.parquet |
| - config_name: t |
| data_files: |
| - split: train |
| path: t/samples.parquet |
| --- |
| # AndroidFlux RL — policy prompts |
|
|
| | Slice | Prompts | Source | |
| |---|---:|---| |
| | t_minus_n | 2,459 | AndroidFlux source trajectories | |
| | successful | 2,623 | AndroidFlux source trajectories | |
| | t_minus_1 | 2,459 | AndroidFlux source trajectories | |
| | t | 2,459 | AndroidFlux source trajectories | |
| | **Subtotal** | **10,000** | | |
| | [ui_genie](ui_genie) | 10,000 | UI-Genie reward-model prompts (5,000-prompt core marked by `in_reduced`) | |
|
|
| The four AndroidFlux slices are drawn from replayed source trajectories. |
| The successful slice contains 53 final-action decisions (2.02%), 53 penultimate-action decisions (2.02%), and 2,517 earlier decisions. These positions refer to the action the policy generates; the last history action is one step earlier. Environment success and canonical train membership are verified. |
|
|
| All 20 canonical unseen-test task types are excluded from them. **Seen-test instances remain in the error-derived slices**, explicitly marked in provenance.jsonl. The two newly added t_minus_n checkpoints are confirmed train. These split rules are about AndroidWorld task types and so do not apply to `ui_genie`, whose contexts come from a different corpus. |
|
|
| Every slice has samples.parquet, context.parquet, images (one `images.parquet`, or `images-NNNNN-of-NNNNN.parquet` shards where the screenshot set is large), samples.json, provenance, and validation/manifests. The Parquet schemas follow [data_from_rm_eval](https://huggingface.co/datasets/Gyubeum/AndroidFlux_RL_Train_Test/tree/main/data_from_rm_eval), adapted to prompt-only system/user conversations with nullable candidate/preference fields. The t slice also has 17,213 unlabelled recovery candidate slots (seven per context), with donor provenance preserved. Per-slice deviations from the reference layout are listed in each `build_manifest.json`. |
|
|
| The `ui_genie` slice is a separate 10,000-prompt release drawn from |
| [`Gyubeum/UI_Genie_128k`](https://huggingface.co/datasets/Gyubeum/UI_Genie_128k) rather |
| than from AndroidFlux trajectories: stratified over history length, reference action type |
| and domain, then filtered by a base-policy rollout whose candidates the reward model had |
| to be able to order. It ships the history screenshots the reward model reads, and the four AndroidFlux slices now also ship their restored history screenshots. See [`ui_genie/SAMPLING.md`](ui_genie/SAMPLING.md), |
| and [`ui_genie/CONTRIBUTING_FORMAT.md`](ui_genie/CONTRIBUTING_FORMAT.md) for the structured |
| format to send new recovery contexts in. |
|
|
| ## AndroidFlux context repair — 2026-09-14 |
|
|
| The original 10,000 AndroidFlux prompt IDs, splits, successful position quotas and t recovery candidates are unchanged. All four slices now embed their full history screenshots, retain clean source-derived or deterministic natural-language action histories alongside the original raw text, and expose recorded next-action references with availability and provenance. References are not verified GT and their coordinates remain source-native. Policy input stays current-image-only; RM readers use context image links or the chronological `rm_image_ids` (last at most five screens). See per-slice READMEs and `repair_validation_summary.json` for exact coverage and limitations. |
|
|
| The additive v2 columns extend (rather than exactly match) the original reference schema. The existing `ui_genie/` release is untouched. `androidflux_tools/export_repaired_interchange.py` produces the documented portable GRPO interchange directly from each repaired Parquet slice, without any original trajectory paths. |
|
|
| ## `is_next_terminate` — 2026-09-24 |
|
|
| `t`, `t_minus_1`, `t_minus_n` and `successful` gain a boolean `is_next_terminate`: the current state already satisfies |
| the goal, so the correct next action is `terminate(status="success")`. `True` is high-precision; `False` means "not |
| confirmed done". Also added: `goal_matches_source_trajectory` (about 15% of the error-derived rows carry a goal that |
| differs from their source trajectory's) and `is_verify_task`. Labelled with `gemini-3.8-flash` (two agreeing passes, a |
| strict gate, cross-checks between neighbouring states) and checked against AndroidWorld outcomes; see |
| [`terminate_annotation.md`](terminate_annotation.md). All other columns and files are unchanged. |
|
|
| ## qwen3vl_replay — 2026-09-24 (use this for Qwen3-VL RL) |
| |
| [`qwen3vl_replay/`](qwen3vl_replay/) rebuilds `t`, `t_minus_1`, `t_minus_n` and `successful` from per-context replay metadata |
| through `Qwen3VLAgent`, exactly as the RL trainer builds its inputs, and ships that metadata. It repairs the goal of the |
| ~15% of rows whose goal came from another episode, drops 172 `t`/`t_minus_1` contexts whose error action the replay does not |
| represent, fixes the replay coordinates of GUI-Owl/MA3 steps, and carries `is_next_terminate` (relabelled where the goal or |
| the `*Verify` evidence window changed). The top-level slices' prompts are superseded but left unchanged; images are shared. |
| Details in [`qwen3vl_replay/README.md`](qwen3vl_replay/README.md). |
| |
