v3 question files renamed to the 2026-10-18 date (contents unchanged); docs updated

.gitattributes

@@ -694,3 +694,6 @@ epic-kitchens/raw/videos/P18/P18_02.MP4 filter=lfs diff=lfs merge=lfs -text
 epic-kitchens/questions/labels_2026-10-12.jsonl filter=lfs diff=lfs merge=lfs -text
 epic-kitchens/questions/train_1012_reftrain.json filter=lfs diff=lfs merge=lfs -text
 epic-kitchens/questions/train_1012.json filter=lfs diff=lfs merge=lfs -text
+epic-kitchens/questions/labels_2026-10-18.jsonl filter=lfs diff=lfs merge=lfs -text
+epic-kitchens/questions/train_2026-10-18.json filter=lfs diff=lfs merge=lfs -text
+epic-kitchens/questions/train_2026-10-18_reftrain.json filter=lfs diff=lfs merge=lfs -text

MD5SUMS.txt

@@ -7,14 +7,14 @@
 1e80108882a36b6fed1461cfbf046897  epic-kitchens/raw/annotations.tar
 38a6d476498d389d80d6c65780b1ffad  epic-kitchens/hand_masks/hand_masks_2026-10-18.tar
 b1afd79090e51cd660141460d9af467c  epic-kitchens/meta/fps_table_2026-10-18.tar
-aad9779a43e6248fa27a6f8c8b4e2e3b  epic-kitchens/questions/LIST_2026-10-12.csv
-d36e6159651c666c67a9f39f332398f4  epic-kitchens/questions/LIST_2026-10-12.md
-3e4c16582c67f2689fc69b9e629f7bce  epic-kitchens/questions/labels_2026-10-12.jsonl
-1b60bcf2188ef510b01ff8305f0ac6fa  epic-kitchens/questions/questions_2026-10-12.tar
-21153945d3f52e8cea999b3018d86895  epic-kitchens/questions/train_1012.json
-c99ce810a38920083e6c9c4b727a3d7f  epic-kitchens/questions/train_1012_reftrain.json
-6dd6140a4ecc84df1f8b651f7455f09d  epic-kitchens/questions/train_1012_reftrain_split.json
-aaf43d53e822df2818797722fe9938ab  epic-kitchens/questions/train_1012_split.json
+aad9779a43e6248fa27a6f8c8b4e2e3b  epic-kitchens/questions/LIST_2026-10-18.csv
+d36e6159651c666c67a9f39f332398f4  epic-kitchens/questions/LIST_2026-10-18.md
+3e4c16582c67f2689fc69b9e629f7bce  epic-kitchens/questions/labels_2026-10-18.jsonl
+ce737696dcfb7826412d4940f7e8d672  epic-kitchens/questions/questions_2026-10-18.tar
+21153945d3f52e8cea999b3018d86895  epic-kitchens/questions/train_2026-10-18.json
+c99ce810a38920083e6c9c4b727a3d7f  epic-kitchens/questions/train_2026-10-18_reftrain.json
+6dd6140a4ecc84df1f8b651f7455f09d  epic-kitchens/questions/train_2026-10-18_reftrain_split.json
+aaf43d53e822df2818797722fe9938ab  epic-kitchens/questions/train_2026-10-18_split.json
 bfc3639a66cebc7757e3250c335c4b91  epic-kitchens/reference_frames/ref512_jpg_2026-10-18.tar
 3f415a69c77e528da9911ca2b42cdd76  epic-kitchens/reference_frames/refgeo-4b_2026-10-18.tar
 db7c1b2fc728c09252446d9fc7224343  epic-kitchens/reference_frames/reftok-4b_2026-10-18.tar

README.md

@@ -10,7 +10,7 @@ The videos are (c) the EPIC-KITCHENS team under CC BY-NC 4.0, and this share kee
 non-commercial research. The original MP4s are not here.
 
 **Current version: 2026-10-18 (v3).** It holds:
-* the 10-12 question set, both tiers (accurate and rough), including the questions built with Gemini;
+* the question set (frozen 2026-10-12), both tiers (accurate and rough), including the questions built with Gemini;
 * one label row per question;
 * the trainable lists;
 * the token cache, 599 videos (the 572 of the previous version + 27 new ones, poses from EPIC-Fields only);
@@ -31,10 +31,10 @@ Every path here is the path under the VoxMem data root (`src/voxmem/paths.py`),
 | path | what |
 |---|---|
 | `epic-kitchens/tokens/qwen3vl-4b_512px/<video>.npz` | 599 files (3.43 TB), one per video. Each holds the Qwen3-VL-4B visual tokens (144 per 512x288 frame), similarity features, 3D points, camera pose, depth and the frame-rule table. Every frame is chosen by the frame rule: each wanted 1 fps frame exactly, else the nearest frame with a camera pose within +-0.5 s, else that second is dropped. sha256 in `epic-kitchens/tokens/tokens_MANIFEST.tsv` and `SHA256SUMS_tokens.txt`. |
-| `epic-kitchens/questions/questions_2026-10-12.tar` | the question set: `2026-10-12/<type>/{accurate,rough}.jsonl` for 25 question types, plus `LIST.{csv,md}` (also loose as `LIST_2026-10-12.{csv,md}`: per type, its counts, whether Gemini was involved, its sources) and `reference_images.json` |
-| `epic-kitchens/questions/labels_2026-10-12.jsonl` | one row per question. The main columns: `tier`, `gemini`, `split`, `coverage`, `n_frames`, `scene_id`, `voxel_tokens`, `est_tokens`, `trainable`, `reason` (the first rule failed) and `trainable_ref_train`. |
-| `epic-kitchens/questions/train_1012.json` + `_split.json` | the 68,465 trainable questions (64,439 train / 4,026 val, on 27 validation videos) |
-| `epic-kitchens/questions/train_1012_reftrain.json` + `_split.json` | the 67,070 of them whose every reference image has 3D (for `--ref_pos train`) |
+| `epic-kitchens/questions/questions_2026-10-18.tar` | the question set: `2026-10-18/<type>/{accurate,rough}.jsonl` for 25 question types, plus `LIST.{csv,md}` (also loose as `LIST_2026-10-18.{csv,md}`: per type, its counts, whether Gemini was involved, its sources) and `reference_images.json` |
+| `epic-kitchens/questions/labels_2026-10-18.jsonl` | one row per question. The main columns: `tier`, `gemini`, `split`, `coverage`, `n_frames`, `scene_id`, `voxel_tokens`, `est_tokens`, `trainable`, `reason` (the first rule failed) and `trainable_ref_train`. |
+| `epic-kitchens/questions/train_2026-10-18.json` + `_split.json` | the 68,465 trainable questions (64,439 train / 4,026 val, on 27 validation videos) |
+| `epic-kitchens/questions/train_2026-10-18_reftrain.json` + `_split.json` | the 67,070 of them whose every reference image has 3D (for `--ref_pos train`) |
 | `epic-kitchens/hand_masks/hand_masks_2026-10-18.tar` | box + palm hand masks, 599 videos each (unpacks to `epic-kitchens/hand_masks/{box,palm}/`). They supersede the loose 572-video `hand_masks/{box,palm}/` files, whose rows are identical. |
 | `epic-kitchens/meta/fps_table_2026-10-18.tar` | `Videos/<P>/{frames,durations}.txt` for the 599 videos: the voxel builder's fps table |
 | `epic-kitchens/reference_frames/reftok-4b_2026-10-18.tar` | 24,727 reference-image token files (144 x 2560, or 192 x 2560 for the two 4:3 videos) |
@@ -42,6 +42,11 @@ Every path here is the path under the VoxMem data root (`src/voxmem/paths.py`),
 | `epic-kitchens/reference_frames/ref512_jpg_2026-10-18.tar` | the 512 px reference images the tokens were made from (needed only to re-bake them) |
 | `epic-kitchens/v3_MANIFEST.tsv` | bytes, md5 and sha256 of every v3 file above except the tokens |
 
+**Renamed on 2026-10-19 so every v3 file carries one date.** The question files were named after the day the questions
+were frozen: `questions_2026-10-12.tar`, `LIST_2026-10-12.{csv,md}`, `labels_2026-10-12.jsonl` and `train_1012[_reftrain][_split].json`.
+Only the names changed. Every loose file is byte-identical, and the tar holds the same 65 files under the folder `2026-10-18/`
+instead of `2026-10-12/`. The old files remain in this repo's history.
+
 ## Earlier versions (kept, not used by v3)
 
 | path | what |

download.sh

@@ -9,7 +9,7 @@ md5sum -c --quiet MD5SUMS.txt --ignore-missing && echo "md5 OK (tars + v3 files
 for t in epic-kitchens/frames/*.tar epic-kitchens/questions/questions.tar epic-kitchens/train/train.tar \
          epic-kitchens/reference_frames/reference_frames.tar epic-kitchens/raw/*.tar \
          epic-kitchens/hand_masks/hand_masks_2026-10-18.tar epic-kitchens/meta/fps_table_2026-10-18.tar \
-         epic-kitchens/questions/questions_2026-10-12.tar; do
+         epic-kitchens/questions/questions_2026-10-18.tar; do
   [ -f "$t" ] || continue; tar -xf "$t" -C "$D" && rm -f "$t"
 done
 echo "unpacked into $D (reftok-4b / refgeo-4b / ref512_jpg tars left in epic-kitchens/reference_frames/: unpack them into <code>/results/cache/hdepic/ and <code>/results/cache/ds/)"

epic-kitchens/questions/LIST_2026-10-18.csv

@@ -0,0 +1,26 @@
+family,type,hd_epic_questions,accurate,rough,gemini_involved,gemini_supplied,other_sources,in_training_set_all_json,source_files
+fine-grained action,fine_grained_action_localization,10000,9689,9689,no,,EK-100 narrations (CPU),yes,data/epic-kitchens/questions/all.json
+fine-grained action,fine_grained_action_recognition,10000,22144,29361,no,,Qwen3.8-27B rewrites narrations from frames; hand-object detections check multi-action,no,results/hdepic_gen/fine_grained_action_recognition/epic/samples_rough.jsonl; results/hdepic_gen/fine_grained_action_recognition/epic_multi/samples_rough.jsonl; results/hdepic_gen/fine_grained_action_recognition/epic_multi_more/samples_rough.jsonl
+fine-grained action,fine_grained_how_recognition,500,1957,3009,no,,"Qwen3.8-27B writes and checks how-clauses from frames (accurate = passes Qwen's self-checks, not human-reviewed)",no,results/hdepic_gen/fine_grained_how_recognition/epic_scale/samples_rough.jsonl
+fine-grained action,fine_grained_why_recognition,500,1993,1993,no,,Qwen3.8-27B picks and checks reasons from frames + narrations (every question passed the checks: no separate rough set),no,results/hdepic_gen/fine_grained_why_recognition/epic_all_v2/samples.jsonl
+3D perception,3d_perception_fixture_location,500,2664,2664,no,,EPIC-Fields poses + VISOR masks (CPU),yes,data/epic-kitchens/questions/all.json
+3D perception,3d_perception_fixture_interaction_counting,300,688,709,no,,"EK-100 narrations + VISOR masks; zero-count answers from the SAM3 kitchen map, Qwen3.8-27B check",yes,data/epic-kitchens/questions/all.json; results/hdepic_gen/3d_perception_fixture_interaction_counting/epic_map/samples_rough.jsonl
+3D perception,3d_perception_object_contents_retrieval,200,3049,4531,no,,"narrations + VISOR + SAM3 boxes, Qwen3.8-27B checks",no,results/hdepic_gen/3d_perception_object_contents_retrieval/epic_v3/samples_rough.jsonl
+3D perception,3d_perception_object_location,500,5366,12047,no,,Qwen3.8-27B boxes and place reads + DA3 depth / EPIC-Fields 3D + kitchen map,no,results/hdepic_gen/3d_perception_object_location/epic_v5_full3d/samples_rough.jsonl
+object motion,object_motion_object_movement_itinerary,500,323,971,no,,"take/put narration chains, Qwen3.8-27B reads, 3D",no,results/hdepic_gen/object_motion_object_movement_itinerary/epic_v1/samples_rough.jsonl
+object motion,object_motion_object_movement_counting,200,321,953,no,,from the itinerary records,no,results/hdepic_gen/object_motion_object_movement_itinerary/epic_v1/movement_counting_rough.jsonl
+object motion,object_motion_stationary_object_localization,200,40,157,no,,from the itinerary records,no,results/hdepic_gen/object_motion_object_movement_itinerary/epic_v1/stationary_localization_rough.jsonl
+gaze,gaze_interaction_anticipation,1000,1243,5797,no,,EK-100 'take' narrations; Qwen3.8-27B finds the head turn (stand-in for gaze),no,results/hdepic_gen/gaze_interaction_anticipation/epic_head/samples_rough.jsonl
+ingredient,ingredient_ingredient_adding_localization,100,2171,2171,no,,EK-100 narrations; Qwen3.8-27B names the dishes (text),yes,data/epic-kitchens/questions/all.json
+ingredient,ingredient_ingredient_retrieval,100,1347,1347,no,,EK-100 narrations (CPU),yes,data/epic-kitchens/questions/all.json
+ingredient,ingredient_ingredients_order,50,308,308,no,,EK-100 narrations (CPU),yes,data/epic-kitchens/questions/all.json
+ingredient,ingredient_ingredient_weight,50,99,99,no,,Qwen3.8-27B reads the scale display; Claude checked every reading; Yushu reviewed all 16 weighings,no,results/hdepic_gen/ingredient_ingredient_weight/questions_v2/short.jsonl
+nutrition,nutrition_nutrition_change,50,74,1039,yes,"ingredient names, add times, amounts (counts / pack / measure / scale; guessed amounts in the rough tier)",SR28 + UK unit weights for nutrition; Qwen3.8-27B typical amounts and agreement checks,no,results/hdepic_gen/nutrition/epic_all_v5/questions_change_rough.jsonl
+nutrition,nutrition_video_nutrition_estimation,50,178,1256,yes,"ingredient names, add times, amounts",SR28 + UK unit weights; Qwen3.8-27B checks; fillers from similar EPIC recipes,no,results/hdepic_gen/nutrition/epic_all_v5/questions_video_rough.jsonl
+recipe,recipe_step_recognition,100,1105,1423,yes,"each video's recipe: names, step texts, step times",Qwen3.8-27B checks the clips,no,results/hdepic_gen/recipe_step_recognition/epic/samples_rough.jsonl
+recipe,recipe_rough_step_localization,100,1105,1423,yes,"each video's recipe: names, step texts, step times",Qwen3.8-27B checks the clips,no,results/hdepic_gen/recipe_rough_step_localization/epic/samples_rough.jsonl
+recipe,recipe_step_localization,50,1105,1423,yes,"each video's recipe: names, step texts, step times",Qwen3.8-27B checks the clips,no,results/hdepic_gen/recipe_step_localization/epic/samples_rough.jsonl
+recipe,recipe_prep_localization,100,694,1108,yes,"each video's recipe: names, step texts, prep times",Qwen3.8-27B checks the clips,no,results/hdepic_gen/recipe_prep_localization/epic/samples_rough.jsonl
+recipe,recipe_multi_step_localization,50,608,1228,yes,"each video's recipe: names, step texts, step and prep times",Qwen3.8-27B checks the clips,no,results/hdepic_gen/recipe_multi_step_localization/epic/samples_rough.jsonl
+recipe,recipe_multi_recipe_recognition,50,439,441,yes,each video's recipe: recipe names (Qwen writes the similar dishes),Qwen3.8-27B checks the clips,no,results/hdepic_gen/recipe_multi_recipe_recognition/epic/samples_rough.jsonl
+recipe,recipe_following_activity_recognition,50,942,1272,yes,each video's recipe: recipe steps and their end times,Qwen3.8-27B checks the clips; labels the activity blocks from 20 s of video + narrations,no,results/hdepic_gen/recipe_following_activity_recognition/epic/samples_rough.jsonl

epic-kitchens/questions/LIST_2026-10-18.md

@@ -0,0 +1,47 @@
+# EPIC-KITCHENS questions in HD-EPIC's 30-type format (2026-10-12)
+
+rough = every question formed (the accurate ones included); accurate = the ones that pass every check. Each type folder holds `rough.jsonl`, `accurate.jsonl` and, where the builder writes one, `quality.jsonl` (why a question is rough). Questions with a box (`<BBOX ...>` in/at `<TIME ...>`) have their image in `epic-kitchens/reference_frames/` (`reference_images.json` maps each question id to `<video>/<frame>_<box>.jpg` with the red box and `<video>/<frame>.jpg` without it; 992 px wide, undistorted). 15 older fixture-counting questions in 8 videos without EPIC-Fields poses have no image. Single-video types only.
+
+| | types | accurate | rough |
+|---|---|---|---|
+| Gemini involved | 9 | 6,250 | 10,613 |
+| no Gemini | 16 | 53,402 | 75,806 |
+| **total** | **25** | **59,652** | **86,419** |
+
+| family | type | HD-EPIC | accurate | rough | Gemini | Gemini supplied | other sources | in all.json |
+|---|---|---|---|---|---|---|---|---|
+| fine-grained action | `fine_grained_action_localization` | 10000 | 9,689 | 9,689 | no |  | EK-100 narrations (CPU) | yes |
+| fine-grained action | `fine_grained_action_recognition` | 10000 | 22,144 | 29,361 | no |  | Qwen3.8-27B rewrites narrations from frames; hand-object detections check multi-action | no |
+| fine-grained action | `fine_grained_how_recognition` | 500 | 1,957 | 3,009 | no |  | Qwen3.8-27B writes and checks how-clauses from frames (accurate = passes Qwen's self-checks, not human-reviewed) | no |
+| fine-grained action | `fine_grained_why_recognition` | 500 | 1,993 | 1,993 | no |  | Qwen3.8-27B picks and checks reasons from frames + narrations (every question passed the checks: no separate rough set) | no |
+| 3D perception | `3d_perception_fixture_location` | 500 | 2,664 | 2,664 | no |  | EPIC-Fields poses + VISOR masks (CPU) | yes |
+| 3D perception | `3d_perception_fixture_interaction_counting` | 300 | 688 | 709 | no |  | EK-100 narrations + VISOR masks; zero-count answers from the SAM3 kitchen map, Qwen3.8-27B check | yes |
+| 3D perception | `3d_perception_object_contents_retrieval` | 200 | 3,049 | 4,531 | no |  | narrations + VISOR + SAM3 boxes, Qwen3.8-27B checks | no |
+| 3D perception | `3d_perception_object_location` | 500 | 5,366 | 12,047 | no |  | Qwen3.8-27B boxes and place reads + DA3 depth / EPIC-Fields 3D + kitchen map | no |
+| object motion | `object_motion_object_movement_itinerary` | 500 | 323 | 971 | no |  | take/put narration chains, Qwen3.8-27B reads, 3D | no |
+| object motion | `object_motion_object_movement_counting` | 200 | 321 | 953 | no |  | from the itinerary records | no |
+| object motion | `object_motion_stationary_object_localization` | 200 | 40 | 157 | no |  | from the itinerary records | no |
+| gaze | `gaze_interaction_anticipation` | 1000 | 1,243 | 5,797 | no |  | EK-100 'take' narrations; Qwen3.8-27B finds the head turn (stand-in for gaze) | no |
+| ingredient | `ingredient_ingredient_adding_localization` | 100 | 2,171 | 2,171 | no |  | EK-100 narrations; Qwen3.8-27B names the dishes (text) | yes |
+| ingredient | `ingredient_ingredient_retrieval` | 100 | 1,347 | 1,347 | no |  | EK-100 narrations (CPU) | yes |
+| ingredient | `ingredient_ingredients_order` | 50 | 308 | 308 | no |  | EK-100 narrations (CPU) | yes |
+| ingredient | `ingredient_ingredient_weight` | 50 | 99 | 99 | no |  | Qwen3.8-27B reads the scale display; Claude checked every reading; Yushu reviewed all 16 weighings | no |
+| nutrition | `nutrition_nutrition_change` | 50 | 74 | 1,039 | yes | ingredient names, add times, amounts (counts / pack / measure / scale; guessed amounts in the rough tier) | SR28 + UK unit weights for nutrition; Qwen3.8-27B typical amounts and agreement checks | no |
+| nutrition | `nutrition_video_nutrition_estimation` | 50 | 178 | 1,256 | yes | ingredient names, add times, amounts | SR28 + UK unit weights; Qwen3.8-27B checks; fillers from similar EPIC recipes | no |
+| recipe | `recipe_step_recognition` | 100 | 1,105 | 1,423 | yes | each video's recipe: names, step texts, step times | Qwen3.8-27B checks the clips | no |
+| recipe | `recipe_rough_step_localization` | 100 | 1,105 | 1,423 | yes | each video's recipe: names, step texts, step times | Qwen3.8-27B checks the clips | no |
+| recipe | `recipe_step_localization` | 50 | 1,105 | 1,423 | yes | each video's recipe: names, step texts, step times | Qwen3.8-27B checks the clips | no |
+| recipe | `recipe_prep_localization` | 100 | 694 | 1,108 | yes | each video's recipe: names, step texts, prep times | Qwen3.8-27B checks the clips | no |
+| recipe | `recipe_multi_step_localization` | 50 | 608 | 1,228 | yes | each video's recipe: names, step texts, step and prep times | Qwen3.8-27B checks the clips | no |
+| recipe | `recipe_multi_recipe_recognition` | 50 | 439 | 441 | yes | each video's recipe: recipe names (Qwen writes the similar dishes) | Qwen3.8-27B checks the clips | no |
+| recipe | `recipe_following_activity_recognition` | 50 | 942 | 1,272 | yes | each video's recipe: recipe steps and their end times | Qwen3.8-27B checks the clips; labels the activity blocks from 20 s of video + narrations | no |
+
+Not staged:
+
+- `ingredient_exact_ingredient_recognition`: built (54) but the weighings span several videos (single-video rule)
+- `ingredient_ingredient_recognition`: multi-video in HD-EPIC: not built
+- `recipe_recipe_recognition`: multi-video in HD-EPIC: not built
+- `nutrition_image_nutrition_estimation`: single-image type: not built
+- `gaze_gaze_estimation`: EPIC has no gaze: not buildable
+
+Action recognition is the newer run (single-action + multi-action), which supersedes the 9,198 in `all.json`. Gemini 3.8 Flash wrote the recipes (458 EPIC videos, about $55) behind the recipe and nutrition types; every Gemini-built set is logged in `results/hdepic_gen/gemini_question_log.jsonl` of the code repository. Qwen = Qwen3.8-27B served with vLLM.

epic-kitchens/questions/labels_2026-10-18.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c08be97677193666869392a745390705ffa89ce3c770b977c5ff407b9a7e1414
+size 42174393

epic-kitchens/questions/questions_2026-10-18.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9d4d9f7bf548387da86c57fa4a64baf85b308e8ae743fc7178d665b74f33e789
+size 636887040

epic-kitchens/questions/train_2026-10-18.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:51c3a287d30b69c958aa4074e3ab095b74116c75b6de7c6737b74816136e77d6
+size 298636787

epic-kitchens/questions/train_2026-10-18_reftrain.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:24e881a1d21c8d54ab05e1f2e510a27087bd0bad7c433885a39e446082b314d5
+size 295310480

epic-kitchens/v3_MANIFEST.tsv

@@ -1,14 +1,14 @@
 path	bytes	md5	sha256
 epic-kitchens/hand_masks/hand_masks_2026-10-18.tar	2520238080	38a6d476498d389d80d6c65780b1ffad	3c8c4e92dd586fb259e1ac6c249b206135a282afe0373528759fc169c4a3fa11
 epic-kitchens/meta/fps_table_2026-10-18.tar	122880	b1afd79090e51cd660141460d9af467c	8d529c49f9ad2fa96e8192426d9fa5ce76fa469dca0012e7f5ddfba79d38c89e
-epic-kitchens/questions/LIST_2026-10-12.csv	5734	aad9779a43e6248fa27a6f8c8b4e2e3b	24badb7dd224095705d265be9f9c6f634f6e039d69f98cd8919394d350c8b233
-epic-kitchens/questions/LIST_2026-10-12.md	5915	d36e6159651c666c67a9f39f332398f4	a95f45d5db6613e8086cb2f51769b7e54a95fe39ff360d1c16cfc9e4bb3ecbda
-epic-kitchens/questions/labels_2026-10-12.jsonl	42174393	3e4c16582c67f2689fc69b9e629f7bce	c08be97677193666869392a745390705ffa89ce3c770b977c5ff407b9a7e1414
-epic-kitchens/questions/questions_2026-10-12.tar	636979200	1b60bcf2188ef510b01ff8305f0ac6fa	8123ebe7f99cc4c54f5fef0e8d7e28e5ad3f8b5d212847f1eb4d15026412c7a6
-epic-kitchens/questions/train_1012.json	298636787	21153945d3f52e8cea999b3018d86895	51c3a287d30b69c958aa4074e3ab095b74116c75b6de7c6737b74816136e77d6
-epic-kitchens/questions/train_1012_reftrain.json	295310480	c99ce810a38920083e6c9c4b727a3d7f	24e881a1d21c8d54ab05e1f2e510a27087bd0bad7c433885a39e446082b314d5
-epic-kitchens/questions/train_1012_reftrain_split.json	1906661	6dd6140a4ecc84df1f8b651f7455f09d	7f562e7e992dce203bb8642e0af0bdf072142cde1d88820360b63fdd172b6b77
-epic-kitchens/questions/train_1012_split.json	1949202	aaf43d53e822df2818797722fe9938ab	346b583f44cb16bf475055bcb3a5ce7a7d6bfd31284916c56e8e8bebbe876e4c
+epic-kitchens/questions/LIST_2026-10-18.csv	5734	aad9779a43e6248fa27a6f8c8b4e2e3b	24badb7dd224095705d265be9f9c6f634f6e039d69f98cd8919394d350c8b233
+epic-kitchens/questions/LIST_2026-10-18.md	5915	d36e6159651c666c67a9f39f332398f4	a95f45d5db6613e8086cb2f51769b7e54a95fe39ff360d1c16cfc9e4bb3ecbda
+epic-kitchens/questions/labels_2026-10-18.jsonl	42174393	3e4c16582c67f2689fc69b9e629f7bce	c08be97677193666869392a745390705ffa89ce3c770b977c5ff407b9a7e1414
+epic-kitchens/questions/questions_2026-10-18.tar	636887040	ce737696dcfb7826412d4940f7e8d672	9d4d9f7bf548387da86c57fa4a64baf85b308e8ae743fc7178d665b74f33e789
+epic-kitchens/questions/train_2026-10-18.json	298636787	21153945d3f52e8cea999b3018d86895	51c3a287d30b69c958aa4074e3ab095b74116c75b6de7c6737b74816136e77d6
+epic-kitchens/questions/train_2026-10-18_reftrain.json	295310480	c99ce810a38920083e6c9c4b727a3d7f	24e881a1d21c8d54ab05e1f2e510a27087bd0bad7c433885a39e446082b314d5
+epic-kitchens/questions/train_2026-10-18_reftrain_split.json	1906661	6dd6140a4ecc84df1f8b651f7455f09d	7f562e7e992dce203bb8642e0af0bdf072142cde1d88820360b63fdd172b6b77
+epic-kitchens/questions/train_2026-10-18_split.json	1949202	aaf43d53e822df2818797722fe9938ab	346b583f44cb16bf475055bcb3a5ce7a7d6bfd31284916c56e8e8bebbe876e4c
 epic-kitchens/reference_frames/ref512_jpg_2026-10-18.tar	917637120	bfc3639a66cebc7757e3250c335c4b91	0f8b4dc389deba2bb39e5c78761575a45aa84d4ebd4e0327945846412de8ee76
 epic-kitchens/reference_frames/refgeo-4b_2026-10-18.tar	184115200	3f415a69c77e528da9911ca2b42cdd76	5a2379b038a871553921af0708edd985a7054e694d205f2701323feb471eb688
 epic-kitchens/reference_frames/reftok-4b_2026-10-18.tar	14258370560	db7c1b2fc728c09252446d9fc7224343	635acf7b8dc86ea18d5e637fe96974509e58655ac6e8d46def72c50b015cf34f