Update OpenCUA numpad multiply trajectory

README.md

@@ -35,7 +35,7 @@ configs:
 
 # cua-lite/OpenCUA
 
-cua-lite preprocessed version of xlangai/AgentNet (OpenCUA). Desktop navigation trajectories from Ubuntu and Windows/Mac environments with pyautogui-style actions converted to CUA-lite tool calls.
+cua-lite preprocessed version of xlangai/AgentNet (OpenCUA). Desktop `use` trajectories from Ubuntu and Windows/Mac environments with pyautogui-style actions converted to CUA-lite tool calls.
 
 ## Origin
 
@@ -53,52 +53,121 @@ ds = load_dataset("cua-lite/OpenCUA")
 ds = load_dataset("cua-lite/OpenCUA", "desktop.use")
 ```
 
-You can also filter by `metadata.platform` / `metadata.task_type` /
-`metadata.others.*` after loading; every row carries a rich `metadata`
-struct (see schema below).
+After loading, parse `metadata` as JSON before filtering by `metadata_kind`,
+`dims`, or `others.*`; every row carries a rich metadata object inside that JSON
+string (see schema below). CUA rows use `dims == [platform, task_type]`.
 
 ## Schema
 
-Each row has these columns:
+Published parquet columns:
 
 | column | type | notes |
 |---|---|---|
 | `images` | list[Image] | embedded PNG/JPEG bytes; HF viewer renders thumbnails |
-| `messages` | list[struct] | OpenAI-style turns with `role` + structured `content` |
-| `metadata` | struct | `{platform, task_type, extra_tool_schemas, valid_actions, others{...}}` |
+| `messages` | string (JSON array) | parse as JSON to OpenAI-style turns with `role`, structured `content`, nested `tool_calls`, and `role:"tool"` results |
+| `metadata` | string (JSON object) | parse as JSON to fields `metadata_kind`, `dims`, `extra_tool_schemas`, CUA-only `valid_actions`, and `others` |
+| `_folded` | string (JSON array, optional) | folded grounding/understanding rows only; authoritative per-instruction `messages` / `metadata` members |
 
 Coordinate values in `messages` are normalized to `[0, 1000]` integers.
+The JSON examples below show the decoded shape, not the raw string cell.
+
+`metadata.extra_tool_schemas[*]` uses the nested Chat Completions function-tool
+declaration shape:
+
+```json
+{
+  "metadata_kind": "cua",
+  "dims": ["desktop", "use"],
+  "extra_tool_schemas": [
+    {
+      "type": "function",
+      "function": {
+        "name": "bash",
+        "description": "Run a shell command.",
+        "parameters": {
+          "type": "object",
+          "properties": {"cmd": {"type": "string"}},
+          "required": ["cmd"]
+        }
+      }
+    }
+  ],
+  "valid_actions": ["click", "type"],
+  "others": {}
+}
+```
+
+`messages[].tool_calls[*]` uses the matching nested invocation shape. Tool
+results pair `tool_calls[].id` with `role:"tool"` `tool_call_id`:
+
+```json
+[
+  {
+    "role": "user",
+    "content": [
+      {"type": "image", "index": 0},
+      {"type": "text", "text": "Click the OK button."}
+    ]
+  },
+  {
+    "role": "assistant",
+    "tool_calls": [
+      {
+        "id": "call_0000",
+        "type": "function",
+        "function": {
+          "name": "computer",
+          "arguments": {
+            "actions": [
+              {"action": "click", "coordinate": [640, 400]},
+              {"action": "type", "text": "hello"}
+            ]
+          }
+        }
+      }
+    ]
+  },
+  {
+    "role": "tool",
+    "tool_call_id": "call_0000",
+    "content": [
+      {"type": "image", "index": 1},
+      {"type": "text", "text": "clicked; typed"}
+    ]
+  }
+]
+```
 
 **Image-dedup (`grounding.*` / `understanding` cohorts).** These cohorts are
 single-image-per-row and many rows share the same screenshot, so to avoid
 re-embedding identical image bytes once per instruction they are stored
 *folded*: one row per unique screenshot (image embedded once), carrying an
 extra **`_folded`** column — a JSON string with the authoritative list of
-`{messages, metadata}` members for that screenshot. The row's top-level
-`messages` is the members concatenated for viewer convenience. `use`
-cohorts are not folded. **Use `lite.data.hf.download` to consume this repo** —
-it unfolds automatically back to one row per instruction; reading the parquet
-directly yields the folded form.
+per-instruction members for that screenshot. Each member's `messages` and
+`metadata` values are the same opaque JSON strings described above. The row's
+top-level `messages` is a JSON string containing the members concatenated for
+viewer convenience. `use` cohorts are not folded. **Use
+`lite.data.hf.download` to consume this repo** — it unfolds automatically back
+to one row per instruction; reading the parquet directly yields the folded form.
 
 ## Layout
 
 ```
-<platform>/<task_type>/<split>/shard-NNNNN-of-NNNNN.parquet                  # single-variant cohort
-<platform>/<task_type>/<split>/<variant>/shard-NNNNN-of-NNNNN.parquet        # multi-variant cohort
+<platform>/<task_type>/<split>/<variant>/shard-NNNNN-of-NNNNN.parquet
 ```
 
-- `platform` ∈ {desktop, mobile, web}
+- `platform` ∈ {desktop, browser, mobile}
 - `task_type` ∈ {understanding, grounding.action, grounding.point, grounding.bbox, use} — used verbatim as the dir component
-- HF config names are `<platform>.<task_type>` (e.g. `mobile.grounding.action`). The agent registry lookup key in code is `<agent>@<platform>@<task_type>` (e.g. `qwen3_vl@mobile@grounding.action`); only this user-facing token uses `.` between platform and task_type, because `@` triggers a 403 on the dataset-viewer's signed image URLs.
+- HF config names are `<platform>.<task_type>` by default (e.g. `mobile.grounding.action`) — UNLESS the dataset was staged with `--config-names`, which sets verbatim, explicitly-chosen config names (see the `configs:` block above for the authoritative list). The agent registry lookup key in code is `<agent>@<platform>@<task_type>` (e.g. `qwen3_vl@mobile@grounding.action`); only this user-facing token uses `.` between platform and task_type, because `@` triggers a 403 on the dataset-viewer's signed image URLs.
 - HF split names stay `train` / `validation` (the `datasets` library blacklists `<>:/\|?*` in split names; everything else is fine in config_name)
-- `validation` is an in-distribution held-out slice (never used in training); `test` is reserved for out-of-distribution benchmark datasets
+- `validation` is an in-distribution held-out slice: no validation **sample** also appears in `train` — content-identical rows (same `images` + same `messages`, differing only in their ids) are co-located into one split, so upstream re-publishing one sample under two ids cannot straddle the split. It is *not* disjoint in **images**: one screenshot legitimately backs many distinct samples, and only whole samples are co-located, so the same picture can appear on both sides. `test` is reserved for out-of-distribution benchmark datasets
 
 ## Stats
 
 | platform | task_type | variant | train | validation |
 |---|---|---|---:|---:|
-| desktop | use | ubuntu | 4,890 | 102 |
-| desktop | use | win_mac | 17,173 | 371 |
+| desktop | use | ubuntu | 4,897 | 102 |
+| desktop | use | win_mac | 16,970 | 367 |
 
 ## Local mirror & SFT export
 
@@ -116,12 +185,12 @@ Rows in the local parquet have `images: list[str]`; bytes are extracted to
 the image store. `lite.train.export.export_sft` consumes the local
 form directly with `--image-root=$CUA_LITE_DATASETS_ROOT`.
 
-- Total unique images: **392,526**
-- Image store size: **186.57 GB**
+- Total unique images: **393,745**
+- Image store size: **186.99 GB**
 
 ## Notes
 
-_(none)_
+split: content-identical rows co-located, then hash_split on metadata.others.id with val_frac=0.02, seed=42, val_cap=2000 per (metadata.dims[0], metadata.dims[1], variant) (an upstream split label, where the source ships one, wins over the hash). A cap-bound carve depends on source iteration order and these parameters do not reproduce it; content co-location may make the final physical validation row count differ from the cap.
 
 ## License & citation
 

desktop/use/train/ubuntu/shard-00371-of-00422.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a48ef86e383ad66f1f31272618a909b5cb6e600d3d868e6a7eb0eab6b9c5f524
-size 247278596
+oid sha256:7fd2c8d451307c8a01552d841bbe6170cf77ebd65b1f095921480c38fd6aacff
+size 271884750

desktop/use/train/ubuntu/shard-00372-of-00422.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ec57c90e21c695a787f0915225875aa110aa012cff7c4670108e63076cea8c35
-size 165299009
+oid sha256:6994624bf47eea6207021a135b2411f680dac3ad9be2b34a36e1a2c72e57b18d
+size 150915102

desktop/use/train/ubuntu/shard-00373-of-00422.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8e5af8a895fb6488f8a8f9446f9c61ab99dd2170e8bcaacfd5cffb546af9274e
-size 134587090
+oid sha256:20bce58a6b491aa40e503974300be0e0cfcfc501316dad66f5fcaf8be23de1d1
+size 146694645

desktop/use/train/ubuntu/shard-00374-of-00422.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d9164668139ae929d4decf76d9b101f04fd788d00ea450c3567c6262dfdb44f3
-size 153906170
+oid sha256:4158014618015cb382a81d1fd5cde4ac8a74a502414bef354d7e2eb841376b2f
+size 165833890

stats.json

@@ -1,13 +1,13 @@
 {
   "rows_in": 0,
-  "rows_out": 22536,
+  "rows_out": 22336,
   "rows_dropped": 0,
-  "unique_images": 392526,
-  "image_store_bytes": 186565347738,
+  "unique_images": 393745,
+  "image_store_bytes": 186988994506,
   "by_partition": {
-    "desktop::use::train::ubuntu": 4890,
-    "desktop::use::train::win_mac": 17173,
+    "desktop::use::train::ubuntu": 4897,
+    "desktop::use::train::win_mac": 16970,
     "desktop::use::validation::ubuntu": 102,
-    "desktop::use::validation::win_mac": 371
+    "desktop::use::validation::win_mac": 367
   }
 }