Add MATT-Bench annotations for Ego4D-M, EPIC-KITCHENS-M, HoloAssist-M

Each dataset lives under its own subdir (ego4d/, epickitchens/, holoassist/)
with xlsx splits as the canonical download format plus parquet mirrors for
the HF dataset viewer. Includes Ego4D's MisEngine reproduction dump
(parquet.xlsx). README rewritten with per-dataset video download
instructions (Ego4D FHO CLI, EPIC-KITCHENS-100 scripts, HoloAssist direct),
schema tables, and an Extended section for HoloAssist-M.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

.gitattributes

@@ -58,3 +58,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+*.xlsx filter=lfs diff=lfs merge=lfs -text

README.md

@@ -15,11 +15,33 @@ tags:
   - semantic-role-labeling
   - ego4d
   - epic-kitchens
+  - holoassist
   - point-of-no-return
   - cvpr2026
 pretty_name: MATT-Bench
 size_categories:
   - 100K<n<1M
+configs:
+  - config_name: ego4d
+    data_files:
+      - split: train
+        path: ego4d/parquet/train.parquet
+      - split: valid
+        path: ego4d/parquet/valid.parquet
+      - split: test
+        path: ego4d/parquet/test.parquet
+  - config_name: epickitchens
+    data_files:
+      - split: train
+        path: epickitchens/parquet/train.parquet
+      - split: validation
+        path: epickitchens/parquet/validation.parquet
+  - config_name: holoassist
+    data_files:
+      - split: train
+        path: holoassist/parquet/train.parquet
+      - split: validation
+        path: holoassist/parquet/validation.parquet
 ---
 
 # Mistake Attribution: Fine-Grained Mistake Understanding in Egocentric Videos
@@ -34,28 +56,151 @@ size_categories:
 
 ---
 
-> **Dataset coming soon.** We are preparing the data for public release. Stay tuned!
-
 ## MATT-Bench Overview
 
-MATT-Bench provides two large-scale benchmarks for **Mistake Attribution (MATT)** — a task that goes beyond binary mistake detection to attribute *what* semantic role was violated, *when* the mistake became irreversible (Point-of-No-Return), and *where* the mistake occurred in the frame.
+MATT-Bench provides large-scale benchmarks for **Mistake Attribution (MATT)** — a task that goes beyond binary mistake detection to attribute *what* semantic role was violated, *when* the mistake became irreversible (Point-of-No-Return), and *where* the mistake occurred in the frame.
 
 The benchmarks are constructed by **MisEngine**, a data engine that automatically creates mistake samples with attribution-rich annotations from existing egocentric action datasets:
 
 | Dataset | Samples | Instruction Texts | Semantic | Temporal | Spatial |
 |---|---|---|---|---|---|
-| **Ego4D-M** | 257,584 | 16,099 | ✓ | ✓ | ✓ |
-| **EPIC-KITCHENS-M** | 221,094 | 12,283 | ✓ | — | — |
+| **Ego4D-M** | 220,800 | 19,467 | ✓ | ✓ | ✓ |
+| **EPIC-KITCHENS-M** | 299,715 | 12,283 | ✓ | — | — |
+
+These are at least **two orders of magnitude larger** than any existing mistake dataset. Instruction-text counts = unique (predicate `V`, argument `ARG1`) pairs.
 
-These are at least **two orders of magnitude larger** than any existing mistake dataset.
+A third source, **HoloAssist-M**, is released alongside as an additional benchmark — see [Extended: HoloAssist-M](#extended-holoassist-m) below.
 
 ## Annotations
 
-Each sample consists of an instruction text and an attempt video, annotated with:
+Each sample consists of an instruction text and an attempt video clip, annotated with:
+
+- **Semantic Attribution**: Which semantic role (predicate `V`, argument `ARG1`) in the instruction is violated in the attempt video.
+- **Temporal Attribution**: The Point-of-No-Return (PNR) frame where the mistake becomes irreversible. Inherited from Ego4D's native PNR annotations — available on **Ego4D-M only**.
+- **Spatial Attribution**: Bounding box localizing the mistake region in the PNR frame. Inherited from Ego4D's native bbox annotations — available on **Ego4D-M only**.
+
+## Repository Layout
+
+```
+MATT-Bench/
+├── ego4d/
+│   ├── train.xlsx, valid.xlsx, test.xlsx   ← primary annotation files (consumed by the MATT codebase)
+│   ├── parquet.xlsx                        ← MisEngine reproduction data (Ego4D narrations with SRL)
+│   └── parquet/                            ← Parquet mirror for the HF dataset viewer
+├── epickitchens/
+│   ├── train.xlsx, validation.xlsx
+│   └── parquet/
+└── holoassist/
+    ├── train.xlsx, validation.xlsx
+    └── parquet/
+```
+
+`.xlsx` is the canonical download format (the MATT codebase reads Excel directly). The `parquet/` mirror powers the HF dataset viewer and `datasets.load_dataset(...)` loaders — both views contain the same rows.
+
+## Downloading MATT-Bench
+
+MATT-Bench has two parts that you obtain separately:
+
+1. **Annotations** — hosted here, download via `hf` or `git clone`.
+2. **Video media** — **not** hosted here. Download from each source dataset using the instructions below. We only mirror the annotations; original videos retain their upstream licenses.
+
+### Annotations (this repo)
+
+```bash
+# Everything
+hf download mistakeattribution/MATT-Bench --repo-type dataset --local-dir MATT-Bench
+
+# Just one source dataset's xlsx files
+hf download mistakeattribution/MATT-Bench --repo-type dataset \
+  --include "ego4d/*.xlsx" --local-dir MATT-Bench
+```
+
+Or via the `datasets` library (reads the parquet mirror):
+
+```python
+from datasets import load_dataset
+ego4d_m = load_dataset("mistakeattribution/MATT-Bench", "ego4d")
+epic_m  = load_dataset("mistakeattribution/MATT-Bench", "epickitchens")
+holo_m  = load_dataset("mistakeattribution/MATT-Bench", "holoassist")
+```
+
+Or read the xlsx directly:
+
+```python
+import pandas as pd
+df = pd.read_excel("MATT-Bench/ego4d/train.xlsx")
+```
+
+### Video media
+
+#### Ego4D (FHO benchmark clips only)
+
+MATT-Bench uses **only the FHO (Forecasting Hands and Objects) benchmark clips** from Ego4D, not the full 3,000-hour dataset.
+
+1. Sign the Ego4D license agreement at <https://ego4d.dev/request/ego4d> (approval ~48h; credentials expire 14 days after approval).
+2. Install the CLI (`pip install ego4d`) and download:
+
+   ```bash
+   ego4d --output_directory="~/ego4d_data" --datasets clips --benchmarks FHO
+   ```
+
+Docs: <https://ego4d-data.org/docs/CLI/>. The `video_uid` and `clip1_uid` fields in our annotations correspond to Ego4D's native video and clip UIDs; `start_frame` / `end_frame` are inherited from Ego4D's FHO annotations.
+
+#### EPIC-KITCHENS-100
 
-- **Semantic Attribution**: Which semantic role (predicate, object) in the instruction is violated in the attempt video
-- **Temporal Attribution**: The Point-of-No-Return (PNR) frame where the mistake becomes irreversible (Ego4D-M)
-- **Spatial Attribution**: Bounding box localizing the mistake region in the PNR frame (Ego4D-M)
+Standard EPIC-KITCHENS-100 download — no access form required.
+
+```bash
+git clone https://github.com/epic-kitchens/epic-kitchens-download-scripts
+cd epic-kitchens-download-scripts
+python epic_downloader.py --rgb-frames    # or --videos
+```
+
+Project page: <https://epic-kitchens.github.io/>. MATT-Bench's `video_id` matches EPIC's participant-video identifier (e.g. `P22_16`); `start_frame` / `end_frame` index the RGB frame sequence.
+
+#### HoloAssist
+
+Standard HoloAssist download — no access form required (CDLAv2 permissive license).
+
+Direct download instructions: <https://holoassist.github.io/>. MATT-Bench's `video_id` matches HoloAssist's video identifier (e.g. `R076-21July-DSLR`).
+
+## Data Schema
+
+### `ego4d/{train,valid,test}.xlsx` — 13 columns
+
+| Column | Description |
+|---|---|
+| `video_uid` | Ego4D video UID (full video) |
+| `start_frame`, `end_frame` | Frame bounds of the attempt clip |
+| `clip1_uid`, `clip1_start_frame`, `clip1_end_frame` | Primary Ego4D clip |
+| `clip2_uid`, `clip2_start_frame`, `clip2_end_frame` | Paired clip for comparison (`Not required` / `-1` when absent) |
+| `V`, `ARG1` | Predicate and argument from the instruction (e.g. `plays`, `lawn tennis`) |
+| `label` | Mistake label |
+| `split` | Split identifier |
+
+### `ego4d/parquet.xlsx` — 29 columns (MisEngine reproduction data)
+
+Ego4D narration-level records with semantic-role labels (`ARG0`, `V`, `ARG1`), frame/time bounds (`start_frame`/`end_frame`/`start_sec`/`end_sec`), clip-relative bounds, and noun/verb embedding vectors. Used to reproduce the MisEngine step that produces the split files above.
+
+### `epickitchens/{train,validation}.xlsx` and `holoassist/{train,validation}.xlsx` — 8 columns
+
+| Column | Description |
+|---|---|
+| `video_id` | Source-dataset video identifier |
+| `start_frame`, `end_frame` | Frame bounds of the attempt clip |
+| `V`, `ARG1` | Predicate and argument from the (possibly corrupted) instruction |
+| `label` | Mistake label |
+| `actual_V`, `actual_ARG1` | Ground-truth predicate/argument as performed |
+
+## Extended: HoloAssist-M
+
+**HoloAssist-M** is an additional MATT benchmark released alongside MATT-Bench. It is **not** part of the main two-dataset evaluation reported in the CVPR 2026 paper; it uses the same MisEngine pipeline applied to the HoloAssist dataset.
+
+| Dataset | Samples | Instruction Texts | Semantic | Temporal | Spatial |
+|---|---|---|---|---|---|
+| **HoloAssist-M** | 562,209 | 1,786 | ✓ | — | — |
+
+Schema matches EPIC-KITCHENS-M (semantic attribution only — HoloAssist does not provide native PNR or bbox annotations).
 
 ## Citation
 
@@ -67,3 +212,28 @@ Each sample consists of an instruction text and an attempt video, annotated with
   year      = {2026},
 }
 ```
+
+Please also cite the source datasets:
+
+```bibtex
+@inproceedings{grauman2022ego4d,
+  title     = {Ego4D: Around the World in 3,000 Hours of Egocentric Video},
+  author    = {Grauman, Kristen and others},
+  booktitle = {CVPR},
+  year      = {2022}
+}
+
+@article{Damen2022RESCALING,
+  title     = {Rescaling Egocentric Vision: Collection, Pipeline and Challenges for EPIC-KITCHENS-100},
+  author    = {Damen, Dima and others},
+  journal   = {IJCV},
+  year      = {2022}
+}
+
+@inproceedings{wang2023holoassist,
+  title     = {HoloAssist: an Egocentric Human Interaction Dataset for Interactive AI Assistants in the Real World},
+  author    = {Wang, Xin and others},
+  booktitle = {ICCV},
+  year      = {2023}
+}
+```

ego4d/parquet.xlsx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:470f94c0b9ec9c62bc37ae8dd1daebce525a38c0b601dc23f2cdc1204784ac7b
+size 19328675

ego4d/parquet/test.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f1ea19c0dd178ea894a9fed0a75e7b702715c99d1f45ed726474dde407eb07ac
+size 651664

ego4d/parquet/train.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:51fc6ad4c1ca9e80c58392e391a6ce16924c521d0f7b5e35047ba1e05faab635
+size 3036071

ego4d/parquet/valid.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2e06e48ebe2ca3b5d3ec9c7159dd3cb423bb826b1a957a8c0b251dc5f410dd33
+size 655928

ego4d/test.xlsx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f107540233d9c792623c78950b4257f41ce73b6698acc9c0845dc5367f72062a
+size 2539715

ego4d/train.xlsx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:37d731094b22f86c1f6751163f899832356fab2324cc8cc85b4248e98ca0e0f7
+size 20375917

ego4d/valid.xlsx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f8460c59490532a0c0446fdc57a3c405214c6edbeb409c6d0cf8fef9b2661568
+size 2541977

epickitchens/parquet/train.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e1deaeec26d434cdcb6e34eca370138f7ee67c360e867e4f3aaeda51a417f0a5
+size 2288186

epickitchens/parquet/validation.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:38e9bd9be789be1cd922c50ea1dc381a882edcff690b14910c1f6c46020c36e0
+size 348141

epickitchens/train.xlsx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:31daef65e02e822112c0823546f337bb376f6b62659306e7f85f371f09534f2b
+size 10767037

epickitchens/validation.xlsx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2239861e5f5d180fd04a44ba94acc422b5fe1fb6c342ba334f9c1866fddb367d
+size 1537253

holoassist/parquet/train.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3d75fd3ad6c67344f65417841e60d0f87af7fcd3692e4a2d81dae3753e18e6b1
+size 3123871

holoassist/parquet/validation.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1db90e922c5d05d9dcfec18f5bdf8e3f5d3ae22b1a708118e23ec4a0af32468c
+size 480023

holoassist/train.xlsx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a2501c0da7fac19c7c8fd748f133e2f971353f1d4d6e1d31930c66357ae37263
+size 20427017

holoassist/validation.xlsx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:88ed419a78cc4a0ab14326cbe9d16eb759c7970833990381aba4ce3a4e3e91d8
+size 2940491