Add natural-language instructions for recorded clicks

index/instructions/<split>.parquet: 663,635 transitions with instructions (551,651 train; 105,703 across the four test splits), each variant with its round-trip result. manifest.json records counts and hashes. The card and docs/schema.md document the table; checksums/sha256sums.txt is updated.

README.md

@@ -617,6 +617,7 @@ to the screen before and after it.
 | Observations | 1,207,368 screenshots in 323,731 scenes |
 | Element instances | 159.7M (132 per screen on average) |
 | Click transitions | 917,211, from 135,731 exploration episodes |
+| Instructions | 663,635 clicks with natural-language instructions |
 | Applications | 19 real Linux desktop applications, several per screen |
 | Appearance | 7 presets, from classic Linux to Windows- and macOS-inspired styles |
 | Resolutions | 7, from 1366×768 to 3840×2160 |
@@ -627,14 +628,14 @@ Splits are made per **scene**, so all frames of a scene stay together. Three tes
 splits hold an attribute out of training **entirely**, which measures
 generalization to desktop configurations never seen in training.
 
-| split | held out | observations | transitions |
-| --- | --- | ---: | ---: |
-| `train` | | 999,494 | 760,850 |
-| `val` | | 11,279 | 8,667 |
-| `test_id` | new scenes with seen attributes | 22,550 | 17,383 |
-| `test_app` | GNOME System Monitor, Pluma, Xarchiver | 43,206 | 32,848 |
-| `test_theme` | the Quartz Night Nord preset | 51,627 | 38,671 |
-| `test_resolution` | 2880×1800 | 79,212 | 58,792 |
+| split | held out | observations | transitions | instructions |
+| --- | --- | ---: | ---: | ---: |
+| `train` | | 999,494 | 760,850 | 551,651 |
+| `val` | | 11,279 | 8,667 | 6,281 |
+| `test_id` | new scenes with seen attributes | 22,550 | 17,383 | 12,589 |
+| `test_app` | GNOME System Monitor, Pluma, Xarchiver | 43,206 | 32,848 | 22,729 |
+| `test_theme` | the Quartz Night Nord preset | 51,627 | 38,671 | 27,175 |
+| `test_resolution` | 2880×1800 | 79,212 | 58,792 | 43,210 |
 
 A held-out application is excluded from every scene that contains it, in any
 frame. See [`docs/splits.md`](docs/splits.md) for how each axis was chosen.
@@ -652,11 +653,13 @@ observation is four members sharing one key:
 | `record.json` | scene metadata (applications, preset, resolution, window stack), eligibility flags, and the action that produced this state |
 
 Alongside the shards, `index/` holds Parquet tables for observations,
-transitions, episodes and scenes: `index/transitions/<split>.parquet` lists every
-recorded click with its before and after observation keys, target element and
-effect. `demo/` holds browsing samples: stratified observations with images
-inline (downscaled to 1024 px) and the `transitions_preview` shards below. The
-field reference is in [`docs/schema.md`](docs/schema.md).
+transitions, instructions, episodes and scenes:
+`index/transitions/<split>.parquet` lists every recorded click with its before
+and after observation keys, target element and effect, and
+`index/instructions/<split>.parquet` gives the clicks their natural-language
+instructions. `demo/` holds browsing samples: stratified observations with
+images inline (downscaled to 1024 px) and the `transitions_preview` shards
+below. The field reference is in [`docs/schema.md`](docs/schema.md).
 
 ### Browsing transitions
 
@@ -675,9 +678,9 @@ types, appearance presets and resolutions. A row reads as one step:
 
 Screenshots are byte-identical copies of the corpus members. Instructions are
 synthesized from each recorded click and its before and after screens; the
-preview shows instructions that an independent grounding model located on the
-before screen alone. It is a sample for browsing: all 917,211 transitions are
-read through `index/transitions` and the shards.
+preview shows instructions that passed the round-trip check described under
+[Instructions](#instructions). It is a sample for browsing: all transitions and
+instructions are read through `index/` and the shards.
 
 ## Loading
 
@@ -698,6 +701,13 @@ transitions = load_dataset(
     data_files="hf://datasets/docling-project/DeskForge-1M/index/transitions/val.parquet",
     split="train",
 )
+
+# Instructions: one row per click, joined to transitions by transition_id.
+instructions = load_dataset(
+    "parquet",
+    data_files="hf://datasets/docling-project/DeskForge-1M/index/instructions/val.parquet",
+    split="train",
+)
 ```
 
 For full passes, read the shards with [`webdataset`](https://github.com/webdataset/webdataset);
@@ -723,22 +733,42 @@ transitions and fetching a single observation by key.
 `transition_train_eligible` selects the action view, which keeps no-op clicks as
 supervision.
 
+## Instructions
+
+663,635 recorded clicks come with natural-language instructions,
+synthesized with Qwen3.6-27B from the click, its target and the screens before
+and after it. An instruction states one coordinate-free goal that can be carried
+out from the before screen, in a `standard` and a more `detailed_contextual`
+style; `primary_instruction` is the standard one where it exists, and
+`referring_expression` describes the target on the before screen. The
+551,651 training instructions are the pool for grounding training, and
+the 105,703 in the four test splits are the grounding evaluation examples.
+
+Each variant also records `roundtrip`: whether an independent grounding model
+(UI-TARS-1.5-7B), given only the unmarked before screen and the instruction,
+points inside the target. It is a strict geometric check, reported rather than
+applied; `primary_roundtrip_pass` selects the subset whose primary instruction
+passed. Counts and source hashes are in `index/instructions/manifest.json`.
+
 ## Annotation quality
 
 Annotations are generated automatically from the accessibility tree and
 reconciled with the screenshot and the window stack. Captures that fail the
 automated audits (element coverage against rendered pixels, blank-widget
 suppression, window ownership and fragment containment) are not included. A
-human audit finds **99.8%** of sampled element annotations correct.
+human audit finds **99.8%** of sampled element annotations correct and
+**97.8%** of sampled instructions sound; instructions are model-generated and
+are not verified row by row.
 
 ## Limitations
 
 All screenshots come from one Linux backend (Xfce): the Windows- and
 macOS-inspired presets reproduce the look of those systems rather than run them.
 Actions are clicks from undirected exploration, not goal-directed
-demonstrations. Because annotations transcribe what is on screen, text drawn by
-applications (for example session paths or live web content in browser scenes)
-is part of the data.
+demonstrations; instructions are written for single clicks after the fact.
+Because annotations transcribe what is on screen, text drawn by applications
+(for example session paths or live web content in browser scenes) is part of the
+data.
 
 ## License
 

checksums/sha256sums.txt

@@ -1150,7 +1150,7 @@ e57fd6a18f21656ebc7a084295cd38683ee89971c89b3f67ee90f3981c69f879  data/val/part-
 5481483852f57ca6f6d5ad07bdebf46ca9ab0982d827ba0106d44e6e9055fef6  data/val/part-00008.tar
 63c97b087a54c529704f566d6619ff7398affca71715ae024084aa48e0ef5e13  data/val/part-00009.tar
 25d4d668aa22fa6e66b3a2078137b2329ae302aacf5291c58dbe847c611e0170  LICENSE
-d15026ecdf8c084e80892c225b7e51d099f6d3712ec72bd809f7f6011747ff23  README.md
+59c5355a3d7d069db511f50db798c92fd39d7dc3caf84c488610b8e45d968469  README.md
 59854984853104df5c353e2f681a15fc7924742f9a2e468c29af248dce45ce03  VERSION
 8c90d39bc3bdc6529f566c68129b495bfc4db79864e5a4c308da710b49fc9678  checksums/shard_stats.parquet
 04b036b3a5449674f0b1f649bfc93c9d6274fb643ee3f54c00201733b0a96aa7  demo/preview-00000-of-00006.parquet
@@ -1165,7 +1165,7 @@ ea01b2c08aaa960bdadc56fe7fcda5a2f3ce48ba3bc6d2aa93b1e97d65675363  demo/transitio
 6179ce5eb9d448e41c898c6f20f260ed477363023f98bd2456c77fc4a68fcbda  demo/transitions/test_theme.tar
 7220d4ad72c3b237811cccd1bdb439924f8dc19195cdbcc59839280fdcb99077  demo/transitions/train.tar
 a217fb98c1f50ce666137fe57f11cc7693bafed9107000b20e0fb83be6aa5ebc  demo/transitions/val.tar
-f4d30a5a1ba3b7cf9f40715c601dd5da328d6e6793d00b69063e3bfa4331469e  docs/schema.md
+0c9f2894d812aeae2b5e03dcc1ca4061b6392afcc50f0eb02de128565756c141  docs/schema.md
 d7eb50809bcf8b048e11fa3a8348bfd5fe84e35621ca6305fd29810e53552ed2  docs/splits.md
 6c652d39ec12c3ada1c06626437fb49934e05f581ff7e3ee140b3f4a02b7bf43  examples/_common.py
 f08b92547613e86ffadfa22decfc96090f0f6919364f94524aa533e1820c77be  examples/get_by_key.py
@@ -1177,6 +1177,13 @@ ea701bf2b1fc18f9c28b97ea0f296e623123051dc66346bfcca5c80a19d66f5e  index/episodes
 1a3aa2a94d138fe19a3d10e8debfb936587217a9c2a6190d32061229dcb9bff6  index/episodes/test_theme.parquet
 14d1458a6d86e8f3ae2782a064a7818e3873786f302e517603ccb3ad0531e7bf  index/episodes/train.parquet
 30c7c147443abb8c52fa71201da43cb8ca6fcfe26fe0ab61d2388df3950a59b3  index/episodes/val.parquet
+174b6b4b031d6d2cc25854b1b8b049293bb6cf2f5eee0e3bde3419b205746737  index/instructions/manifest.json
+98d2c230bfc73b94ed5b9d832dfd58353f9a7ffb7509388b08e41c9433b97d94  index/instructions/test_app.parquet
+5333296f6c73f661fc2211a887912c002993a67febd64296c5f7cc646bfca877  index/instructions/test_id.parquet
+95217e12416731af0c92a3176f36029915bc4755fab4a1d4a9454d705b9219e6  index/instructions/test_resolution.parquet
+048aaed4da4aa6ee856cdc62561761a4d9c0a7c695f5179b117e1c1960c5dca2  index/instructions/test_theme.parquet
+a749d691cff27154b08c53de18584cadda24e290a76fc9ab938d474bd1669fd1  index/instructions/train.parquet
+b0ae7f1092b95c64d66eac7b9017864297b81b17d5b4586c787d3a199aa37620  index/instructions/val.parquet
 9701bf69b26dc6a1c6cd8743dec93b51d7611dab3db23feca308ec94af905a90  index/observations/test_app.parquet
 fd25ec025707ac484aaa110525f3078cc20a28659221e64b74f5378b2e5f4dbe  index/observations/test_id.parquet
 b1bc202bc2fe24469e70a4bc593fc3bbd6d8d931e488cf73d4a7dcfdb98c6ae3  index/observations/test_resolution.parquet

docs/schema.md

@@ -69,6 +69,7 @@ observation that an action led to — `action_into_this_state` and
 | `index/scenes.parquet` | scene | `split`, `split_source`, `apps`, `theme`, `resolution` |
 | `index/observations/<split>.parquet` | observation | `tar_path`, `*_member`, `occluded_ratio`, `n_elements`, `n_windows`, the flags |
 | `index/transitions/<split>.parquet` | transition | `before_key`, `after_key`, action fields, `effect`, `exclusion_reasons` |
+| `index/instructions/<split>.parquet` | transition with an instruction | `primary_instruction`, `instruction_variants`, `referring_expression`, `primary_roundtrip_pass` |
 | `index/episodes/<split>.parquet` | episode | `observation_keys`, `transition_ids`, `episode_status` |
 | `index/shard_members.parquet` | tar member | `byte_offset`, `byte_size`, `sha256` |
 | `checksums/shard_stats.parquet` | tar | `bytes`, `sha256`, `observations` |
@@ -100,6 +101,27 @@ Coordinates: `action_point_px` is authoritative. `action_point_norm_1000` and
 otherwise names every rule the transition failed, so a count can always be
 accounted for.
 
+## Instructions
+
+`index/instructions/<split>.parquet` has one row per transition with an
+instruction, joined to the corpus by `transition_id`, `before_key` and
+`after_key`:
+
+| column | content |
+| --- | --- |
+| `transition_id`, `split`, `before_key`, `after_key`, `target_uid` | the transition and its target element |
+| `action_type`, `action_point_px`, `action_target_bbox_px` | the recorded click, as in `index/transitions` |
+| `primary_style`, `primary_instruction` | the main instruction: `standard` where it exists, else `detailed_contextual` |
+| `instruction_variants` | every kept phrasing: `style`, `text` and `roundtrip` (`pass` or `fail`) |
+| `referring_expression` | how the target is identified on the before screen |
+| `primary_roundtrip_pass` | whether the primary instruction passed the round trip |
+
+Instructions are coordinate-free: they name a goal, never a position. The round
+trip gives an independent grounding model (UI-TARS-1.5-7B) only the unmarked
+before screen and one instruction, and passes when its point falls inside the
+target's visible fragments. `manifest.json` records per-split counts, file
+hashes and the source hashes.
+
 ## `demo/preview.parquet`
 
 A stratified sample for the Hub's Dataset Viewer, one row per observation with

index/instructions/manifest.json

@@ -0,0 +1,73 @@
+{
+  "description": "Natural-language instructions for recorded clicks: one row per transition, joined to the corpus by transition_id, before_key and after_key. Model-generated; not verified row by row.",
+  "instruction_schema_version": "semantic-v2.1-candidate",
+  "generator": "Qwen3.6-27B",
+  "styles": [
+    "standard",
+    "detailed_contextual"
+  ],
+  "primary_style_selection": "first_available_in_included_styles_order",
+  "roundtrip": {
+    "verifier": "ByteDance-Seed/UI-TARS-1.5-7B",
+    "verifier_revision": "683d002dd99d8f95104d31e70391a39348857f4e",
+    "criterion": "given only the unmarked before screen and the instruction, the verifier's point falls inside the target's visible fragments",
+    "variants_checked": 1221534,
+    "variants_passed": 768766
+  },
+  "splits": {
+    "train": {
+      "rows": 551651,
+      "variants": 1014639,
+      "variants_passed": 645390,
+      "primary_passed": 335582,
+      "sha256": "a749d691cff27154b08c53de18584cadda24e290a76fc9ab938d474bd1669fd1"
+    },
+    "val": {
+      "rows": 6281,
+      "variants": 11552,
+      "variants_passed": 7256,
+      "primary_passed": 3772,
+      "sha256": "b0ae7f1092b95c64d66eac7b9017864297b81b17d5b4586c787d3a199aa37620"
+    },
+    "test_id": {
+      "rows": 12589,
+      "variants": 23289,
+      "variants_passed": 14840,
+      "primary_passed": 7716,
+      "sha256": "5333296f6c73f661fc2211a887912c002993a67febd64296c5f7cc646bfca877"
+    },
+    "test_app": {
+      "rows": 22729,
+      "variants": 41727,
+      "variants_passed": 25777,
+      "primary_passed": 13355,
+      "sha256": "98d2c230bfc73b94ed5b9d832dfd58353f9a7ffb7509388b08e41c9433b97d94"
+    },
+    "test_theme": {
+      "rows": 27175,
+      "variants": 50384,
+      "variants_passed": 31088,
+      "primary_passed": 15883,
+      "sha256": "048aaed4da4aa6ee856cdc62561761a4d9c0a7c695f5179b117e1c1960c5dca2"
+    },
+    "test_resolution": {
+      "rows": 43210,
+      "variants": 79943,
+      "variants_passed": 44415,
+      "primary_passed": 22725,
+      "sha256": "95217e12416731af0c92a3176f36029915bc4755fab4a1d4a9454d705b9219e6"
+    }
+  },
+  "source": {
+    "instruction_set": "1.1.0-rc2",
+    "files_sha256": {
+      "train": "88c44c4cfdc5e94c62e5a86893ca76d93167c6647b07c206b35ff1e96f5240e6",
+      "val": "0190f978843b179da32ce052ae5186b8786b666b4542bd82d7558b0a2a24e187",
+      "test_id": "4368c7dd321db4b8792454ea4345e2fb58d4275a5d7c354af5955dc805bc7048",
+      "test_app": "c2418f2ef3cb4e9440a97b37fa125957f58f69ad298897af967c47afebea6ba8",
+      "test_theme": "8b7e1d98973c09c5ac4cb65295861d0b3a0f3f94ff833eced0c80b1e2d0a1562",
+      "test_resolution": "944b28ef211a1f417656724dbd634f1ff008f2621af5da6311389d9e90ba66c7"
+    },
+    "roundtrip_validation_sha256": "63b9a3a75cd64be9693d5be7c0b2dc655d6562f6f9eea594bcc1db68dcccbb15"
+  }
+}

index/instructions/test_app.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:98d2c230bfc73b94ed5b9d832dfd58353f9a7ffb7509388b08e41c9433b97d94
+size 1969440

index/instructions/test_id.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5333296f6c73f661fc2211a887912c002993a67febd64296c5f7cc646bfca877
+size 1098917

index/instructions/test_resolution.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:95217e12416731af0c92a3176f36029915bc4755fab4a1d4a9454d705b9219e6
+size 3857159

index/instructions/test_theme.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:048aaed4da4aa6ee856cdc62561761a4d9c0a7c695f5179b117e1c1960c5dca2
+size 2375638

index/instructions/train.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a749d691cff27154b08c53de18584cadda24e290a76fc9ab938d474bd1669fd1
+size 53295498

index/instructions/val.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b0ae7f1092b95c64d66eac7b9017864297b81b17d5b4586c787d3a199aa37620
+size 576950