Add README.md for EgoTempo manifest v1

README.md

@@ -1,109 +1,36 @@
 ---
-dataset_info:
-  features:
-  - name: schema_version
-    dtype: large_string
-  - name: dataset_id
-    dtype: large_string
-  - name: benchmark
-    dtype: large_string
-  - name: split
-    dtype: large_string
-  - name: base_dataset
-    dtype: large_string
-  - name: question_id
-    dtype: large_string
-  - name: annotation_id
-    dtype: large_string
-  - name: query_index
-    dtype: int64
-  - name: category
-    dtype: large_string
-  - name: level
-    dtype: 'null'
-  - name: task_main_category
-    dtype: large_string
-  - name: task_subcategory
-    dtype: large_string
-  - name: task_type_name
-    dtype: large_string
-  - name: question
-    dtype: large_string
-  - name: choices
-    large_list: 'null'
-  - name: answer_index
-    dtype: 'null'
-  - name: answer_text
-    dtype: large_string
-  - name: answer_key_status
-    dtype: large_string
-  - name: source_ids
-    large_list: large_string
-  - name: source_group_key
-    dtype: large_string
-  - name: primary_asset_id
-    dtype: large_string
-  - name: asset_ids
-    large_list: large_string
-  - name: asset_count
-    dtype: int64
-  - name: evidence_count
-    dtype: int64
-  - name: video_relpath
-    dtype: large_string
-  - name: video_start_s
-    dtype: float64
-  - name: video_end_s
-    dtype: float64
-  - name: query_time_s
-    dtype: float64
-  - name: raw_query_time_s
-    dtype: 'null'
-  - name: effective_query_time_s
-    dtype: 'null'
-  - name: causal_input_start_s
-    dtype: float64
-  - name: causal_input_end_s
-    dtype: float64
-  - name: query_time_source
-    dtype: large_string
-  - name: query_time_adjusted
-    dtype: bool
-  - name: query_duration_delta_s
-    dtype: 'null'
-  - name: evaluation_context
-    dtype: large_string
-  - name: evidence_intervals
-    large_list: 'null'
-  - name: evidence_intervals_causal
-    large_list: 'null'
-  - name: required_modalities
-    large_list: large_string
-  - name: modalities
-    large_list: large_string
-  - name: local_video_available
-    dtype: bool
-  - name: coverage_status
-    dtype: large_string
-  - name: image_option_count
-    dtype: int64
-  - name: future_frame_allowed
-    dtype: bool
-  - name: native_fps
-    dtype: 'null'
-  - name: source_license
-    dtype: large_string
-  - name: source_url
-    dtype: large_string
-  splits:
-  - name: test
-    num_bytes: 642968
-    num_examples: 500
-  download_size: 192981
-  dataset_size: 642968
-configs:
-- config_name: default
-  data_files:
-  - split: test
-    path: data/test-*
+pretty_name: EgoTempo video-only common manifest
+license: cc-by-4.0
+tags:
+- egocentric-video
+- temporal-reasoning
+- benchmark-manifest
 ---
+
+# EgoTempo
+
+500 public open-ended test questions from the official EgoTempo release. This is an annotation-only
+common-schema manifest: source Ego4D videos are not included and `local_video_available` is false.
+No IMU fields are present because this release does not provide an IMU stream.
+
+The official clip identifier encodes the Ego4D source UID and clip start/end. The manifest uses the
+clip interval as a full-clip causal input and sets `evaluation_context` to
+`full_clip_causal_at_clip_end`. There is no official per-question query timestamp or evidence
+interval in this release, so the query endpoint is explicitly derived from the clip end; this is
+not an oracle evidence claim and future frames beyond the clip are not allowed.
+
+The public answer is open-ended text (`answer_text`), not a multiple-choice index. Evaluation follows
+the official open-ended/reference-answer protocol rather than exact MC accuracy.
+
+## Canonical files
+
+- `benchmark.parquet`: nested-schema canonical artifact.
+- `benchmark.jsonl`: line-oriented canonical artifact.
+- `summary.json`: counts, provenance, and license boundary.
+
+## Provenance and redistribution
+
+Annotation source: https://github.com/google-research-datasets/egotempo
+Release commit: `7022ba77b4d89f51cf34e499767995ccd5c90c7a`
+License: CC-BY-4.0 for the EgoTempo release. Ego4D source video remains external and must be
+obtained through the official Ego4D access/license process; it is not redistributed here.