{ "title": "Ropedia Xperience-10M Glossary", "status": "published", "purpose": "Define reader-facing project terms and adjacent technical field terms for embodied AI, egocentric multimodal data, spatial intelligence, world models, VLA/policy learning, evaluation, and public artifact reading.", "categories": [ { "id": "dataset_scope", "label": "Dataset and scope", "description": "Public data boundaries, evidence lines, and how each result family should be read." }, { "id": "files_features", "label": "Files and features", "description": "Raw sample files, windows, feature manifests, and public-safe derivatives." }, { "id": "multimodal_sensing", "label": "Multimodal sensing", "description": "Video, audio, depth, IMU, motion capture, calibration, and synchronization terms." }, { "id": "spatial_geometry", "label": "Spatial geometry", "description": "Camera pose, SLAM, coordinate frames, point clouds, 3D reconstruction, and spatial grounding." }, { "id": "temporal_world_models", "label": "Temporal and world models", "description": "Future prediction, rollouts, forward dynamics, long-horizon forecasting, and temporal leakage." }, { "id": "robotics_vla", "label": "Robotics and VLA", "description": "Vision-language-action, policies, action chunks, imitation learning, contact, and dexterity." }, { "id": "tasks_metrics", "label": "Tasks and metrics", "description": "Task contracts, scored records, direct scores, compact proxies, and audits." }, { "id": "training_eval", "label": "Training and evaluation", "description": "Splits, held-out evaluation, metric types, prompt/schema checks, adapters, and distributed training." }, { "id": "models_runs", "label": "Models and runs", "description": "Baseline families, Qwen3-Omni, Cosmos3, LoRA adapters, and full-parameter gates." }, { "id": "public_surfaces", "label": "Public surfaces", "description": "GitHub, website, Hugging Face repos, parity checks, and package validation." } ], "entries": [ { "term": "Evidence line", "category": "dataset_scope", "plain_meaning": "A reading lane for a group of results.", "project_usage": "Line 1 is one public sample episode; Line 2 is selected-128 held-out comparison.", "do_not_confuse_with": "Qwen run versions v1-v6, which are model-run lineage.", "primary_files": [ "TWO_EVIDENCE_LINES.md", "docs/data/two_evidence_lines.json" ] }, { "term": "Official gated data", "category": "dataset_scope", "plain_meaning": "Upstream files that require official dataset access.", "project_usage": "Raw Xperience-10M MP4/HDF5/RRD files and full source directories remain outside the public repo.", "do_not_confuse_with": "Public-safe metrics, derived features, figures, and manifests.", "primary_files": [ "DATA_NOTICE.md", "REPRODUCIBILITY.md" ] }, { "term": "Public sample episode", "category": "dataset_scope", "plain_meaning": "One officially available sample episode.", "project_usage": "The fully inspectable Line 1 unit used for raw-file browsing, 20-frame windows, task construction, and single-episode baselines.", "do_not_confuse_with": "The selected-128 comparison rows.", "primary_files": [ "docs/data/raw_sample_files.json", "docs/single_episode_explorer.html" ] }, { "term": "Selected 128 episodes", "category": "dataset_scope", "plain_meaning": "A public-safe selected subset of official gated episode paths.", "project_usage": "Line 2 uses derived windows/features and keeps links back to official episode ids and gated source paths.", "do_not_confuse_with": "Redistributed raw MP4/HDF5/RRD data.", "primary_files": [ "XPERIENCE10M_128_EPISODE_FEATURE_INDEX.md", "docs/data/xperience10m_128_episode_feature_index.json" ] }, { "term": "Xperience-10M", "category": "dataset_scope", "plain_meaning": "The upstream embodied human-interaction dataset.", "project_usage": "Source dataset behind the public sample, selected-128 features, task suite, and model diagnostics.", "do_not_confuse_with": "This repo, which only redistributes public-safe derived artifacts.", "primary_files": [ "XPERIENCE10M_DATASET_CARD_ALIGNMENT.md", "docs/data/xperience10m_dataset_card_alignment.json" ] }, { "term": "20-frame window", "category": "files_features", "plain_meaning": "A fixed short clip slice.", "project_usage": "The sample episode is converted into aligned 20-frame units for features, labels, and many task heads.", "do_not_confuse_with": "A full episode or arbitrary video segment.", "primary_files": [ "results/episode_task_suite/windows.csv", "EVALUATION_PROTOCOL.md" ] }, { "term": "annotation.hdf5", "category": "files_features", "plain_meaning": "Upstream annotation container for the sample.", "project_usage": "Contains original labels/metadata; some public derived files expose processed features instead of every raw text field.", "do_not_confuse_with": "Task result summaries.", "primary_files": [ "docs/data/raw_sample_files.json" ] }, { "term": "Episode", "category": "files_features", "plain_meaning": "One recorded interaction sequence.", "project_usage": "The basic source unit behind windows, labels, and train/val/test splits.", "do_not_confuse_with": "A 20-frame window.", "primary_files": [ "docs/data/raw_sample_files.json", "docs/data/xperience10m_128_episode_feature_index.json" ] }, { "term": "Feature manifest", "category": "files_features", "plain_meaning": "A map from model-input columns to source modalities.", "project_usage": "Explains feature groups and dimensions for the sample task suite.", "do_not_confuse_with": "The raw annotation file.", "primary_files": [ "results/episode_task_suite/feature_manifest.json" ] }, { "term": "Interaction text", "category": "files_features", "plain_meaning": "Natural-language interaction/caption content.", "project_usage": "Used by task 15 and some derived text features; public matrices record direct or compact-proxy status.", "do_not_confuse_with": "Numeric action ids or subtask ids.", "primary_files": [ "TASK_SUITE_20.md", "docs/data/task_method_20_result_matrix.json" ] }, { "term": "Modality", "category": "files_features", "plain_meaning": "A type of signal.", "project_usage": "Video, audio, depth, pose/SLAM, motion capture, inertial, calibration, and language-derived signals.", "do_not_confuse_with": "A task target.", "primary_files": [ "docs/data/modality_atlas.json", "results/episode_task_suite/feature_manifest.json" ] }, { "term": "Raw sample file map", "category": "files_features", "plain_meaning": "A human-readable inventory of the sample episode files.", "project_usage": "Explains videos, annotations, calibration, motion, and derived previews.", "do_not_confuse_with": "A training manifest.", "primary_files": [ "docs/data/raw_sample_files.json" ] }, { "term": "visualization.rrd", "category": "files_features", "plain_meaning": "Rerun viewer recording for visual inspection.", "project_usage": "Can be downloaded from the official sample dataset and opened in Rerun 0.29.0 to inspect the sample episode. It is not used for published training or metric rows.", "do_not_confuse_with": "MP4 video streams or model inputs.", "primary_files": [ "docs/data/raw_sample_files.json", "REPRODUCIBILITY.md" ] }, { "term": "Window stride", "category": "files_features", "plain_meaning": "The frame step between neighboring windows.", "project_usage": "Creates overlapping examples while preserving chronological order and leakage controls.", "do_not_confuse_with": "Video frame rate.", "primary_files": [ "EVALUATION_PROTOCOL.md" ] }, { "term": "Audio waveform", "category": "multimodal_sensing", "plain_meaning": "A time-series pressure signal from sound.", "project_usage": "The audio ablation measures whether embedded audio helps selected task contracts.", "do_not_confuse_with": "Language captions or text labels.", "primary_files": [ "docs/data/audio_ablation_summary.json" ] }, { "term": "Calibration", "category": "multimodal_sensing", "plain_meaning": "Parameters that relate sensors to each other and to physical space.", "project_usage": "Needed to interpret camera streams, depth, pose, and synchronized multimodal features together.", "do_not_confuse_with": "A model training hyperparameter.", "primary_files": [ "docs/data/raw_sample_files.json" ] }, { "term": "Camera extrinsics", "category": "multimodal_sensing", "plain_meaning": "A camera position and orientation relative to another coordinate frame.", "project_usage": "Connects different camera streams and world coordinates.", "do_not_confuse_with": "Camera intrinsics.", "primary_files": [ "docs/data/raw_sample_files.json" ] }, { "term": "Camera intrinsics", "category": "multimodal_sensing", "plain_meaning": "Internal camera parameters such as focal length and distortion.", "project_usage": "Explain how image pixels project to rays for geometry tasks.", "do_not_confuse_with": "Camera extrinsics.", "primary_files": [ "docs/data/raw_sample_files.json" ] }, { "term": "Depth map", "category": "multimodal_sensing", "plain_meaning": "A per-pixel estimate of distance from the camera.", "project_usage": "Depth-derived signals support spatial and geometry-oriented tasks.", "do_not_confuse_with": "RGB brightness or semantic segmentation.", "primary_files": [ "docs/data/modality_atlas.json" ] }, { "term": "Egocentric video", "category": "multimodal_sensing", "plain_meaning": "Video captured from a first-person or body-mounted viewpoint.", "project_usage": "The sample streams are egocentric views of human interaction and are the visual basis for many tasks.", "do_not_confuse_with": "Third-person robot-camera footage.", "primary_files": [ "docs/data/raw_sample_files.json" ] }, { "term": "Fisheye camera", "category": "multimodal_sensing", "plain_meaning": "A wide-angle camera with strong lens distortion.", "project_usage": "Multiple fisheye MP4 streams give broad room coverage but need calibration-aware interpretation.", "do_not_confuse_with": "A rectilinear pinhole camera image.", "primary_files": [ "docs/data/raw_sample_files.json" ] }, { "term": "IMU", "category": "multimodal_sensing", "plain_meaning": "An inertial measurement unit with accelerometer and gyroscope signals.", "project_usage": "Supports motion, temporal, and sensor-bridging tasks.", "do_not_confuse_with": "Motion capture skeleton data.", "primary_files": [ "docs/data/modality_atlas.json" ] }, { "term": "Metric depth", "category": "multimodal_sensing", "plain_meaning": "Depth expressed in physical units rather than arbitrary relative scale.", "project_usage": "Useful for distance-sensitive spatial reasoning and reconstruction targets.", "do_not_confuse_with": "Relative monocular depth.", "primary_files": [ "docs/data/modality_atlas.json" ] }, { "term": "Motion capture", "category": "multimodal_sensing", "plain_meaning": "A system that records body or hand motion over time.", "project_usage": "Provides hand/body motion evidence when exposed through public-safe derived features.", "do_not_confuse_with": "Video-only pose estimation.", "primary_files": [ "docs/data/modality_atlas.json" ] }, { "term": "RGB frame", "category": "multimodal_sensing", "plain_meaning": "A color image frame from a video stream.", "project_usage": "Used for visual statistics, previews, and many model inputs.", "do_not_confuse_with": "Depth values or point-cloud coordinates.", "primary_files": [ "results/episode_task_suite/feature_manifest.json" ] }, { "term": "Sensor alignment", "category": "multimodal_sensing", "plain_meaning": "Putting different sensor streams into a shared temporal or spatial reference.", "project_usage": "Used to make video, audio, pose, depth, IMU, and mocap usable in the same task input.", "do_not_confuse_with": "Model ensembling.", "primary_files": [ "docs/data/modality_atlas.json" ] }, { "term": "Stereo camera", "category": "multimodal_sensing", "plain_meaning": "A paired-camera setup that supports depth or geometry estimation.", "project_usage": "The sample browser exposes stereo streams as part of the visual modality set.", "do_not_confuse_with": "Single-view RGB video.", "primary_files": [ "docs/data/raw_sample_files.json" ] }, { "term": "Timestamp synchronization", "category": "multimodal_sensing", "plain_meaning": "Aligning sensor samples by time.", "project_usage": "The task suite assumes aligned windows across modalities so labels and features refer to the same moment.", "do_not_confuse_with": "Randomly joining files with similar names.", "primary_files": [ "EVALUATION_PROTOCOL.md" ] }, { "term": "3D reconstruction", "category": "spatial_geometry", "plain_meaning": "Recovering 3D scene structure from sensor data.", "project_usage": "One core spatial-intelligence direction for Xperience-style data.", "do_not_confuse_with": "Next-action classification.", "primary_files": [ "docs/data/three_foundation_pipelines.json" ] }, { "term": "Affordance", "category": "spatial_geometry", "plain_meaning": "An action possibility offered by an object or scene.", "project_usage": "Relevant when moving from observed human interaction to robot-action or VLA tasks.", "do_not_confuse_with": "A detected object category alone.", "primary_files": [ "docs/data/three_foundation_pipelines.json" ] }, { "term": "Camera pose", "category": "spatial_geometry", "plain_meaning": "The camera position and orientation at a time step.", "project_usage": "Supports spatial-intelligence tasks, view synchronization, and geometry diagnostics.", "do_not_confuse_with": "The human body pose.", "primary_files": [ "docs/data/modality_atlas.json" ] }, { "term": "Coordinate frame", "category": "spatial_geometry", "plain_meaning": "A reference system for positions and orientations.", "project_usage": "Needed when comparing camera, body, object, and world measurements.", "do_not_confuse_with": "A video frame.", "primary_files": [ "EVALUATION_PROTOCOL.md" ] }, { "term": "Object-centric representation", "category": "spatial_geometry", "plain_meaning": "A representation organized around objects and their relations.", "project_usage": "Useful for object relevance, object-set forecast, and action-object relation tasks.", "do_not_confuse_with": "A flat feature vector without object identity.", "primary_files": [ "docs/data/task_suite_20.json" ] }, { "term": "Odometry", "category": "spatial_geometry", "plain_meaning": "Motion estimated from sensor changes over time.", "project_usage": "A relevant spatial term for ego-motion and camera-pose reasoning.", "do_not_confuse_with": "Ground-truth motion capture.", "primary_files": [ "docs/data/modality_atlas.json" ] }, { "term": "Point cloud", "category": "spatial_geometry", "plain_meaning": "A set of 3D points representing scene structure.", "project_usage": "A likely target or intermediate representation for spatial-intelligence extensions.", "do_not_confuse_with": "A 2D image grid.", "primary_files": [ "docs/data/three_foundation_pipelines.json" ] }, { "term": "SLAM", "category": "spatial_geometry", "plain_meaning": "Simultaneous localization and mapping.", "project_usage": "A field term for estimating camera motion and scene structure from sensor observations.", "do_not_confuse_with": "A task label or action class.", "primary_files": [ "docs/data/modality_atlas.json" ] }, { "term": "Spatial grounding", "category": "spatial_geometry", "plain_meaning": "Linking language or labels to locations, objects, or geometry.", "project_usage": "Connects language grounding tasks with 3D/spatial reasoning.", "do_not_confuse_with": "General text classification.", "primary_files": [ "docs/data/research_directions.json" ] }, { "term": "Trajectory", "category": "spatial_geometry", "plain_meaning": "A sequence of positions over time.", "project_usage": "Used for hand motion, camera motion, and future-path tasks.", "do_not_confuse_with": "A single coordinate or label.", "primary_files": [ "TASK_SUITE_20.md" ] }, { "term": "Action forecasting", "category": "temporal_world_models", "plain_meaning": "Predicting a future action before it happens.", "project_usage": "Covered by next-action and long-horizon task contracts.", "do_not_confuse_with": "Recognizing the current action only.", "primary_files": [ "docs/data/task_suite_20.json" ] }, { "term": "Autoregressive prediction", "category": "temporal_world_models", "plain_meaning": "Generating each future token, state, or frame conditioned on prior outputs.", "project_usage": "Relevant for model branches that produce structured JSON or temporal predictions.", "do_not_confuse_with": "A one-shot classifier.", "primary_files": [ "docs/data/foundation_model_plan.json" ] }, { "term": "Forward dynamics", "category": "temporal_world_models", "plain_meaning": "Predicting the next state from the current state and action/context.", "project_usage": "The Cosmos3-Super LoRA branch uses a forward-dynamics-style diagnostic contract.", "do_not_confuse_with": "Reverse inference from result back to cause.", "primary_files": [ "docs/data/omni_model_comparison.json" ] }, { "term": "Latent state", "category": "temporal_world_models", "plain_meaning": "A hidden representation that summarizes observed context.", "project_usage": "Useful for future foundation-model and world-model training plans.", "do_not_confuse_with": "A visible annotation column.", "primary_files": [ "docs/data/foundation_model_plan.json" ] }, { "term": "Long-horizon prediction", "category": "temporal_world_models", "plain_meaning": "Predicting outcomes several seconds or steps ahead.", "project_usage": "Tasks 13 and 14 test longer temporal context beyond immediate recognition.", "do_not_confuse_with": "Single-frame classification.", "primary_files": [ "docs/data/task_suite_20.json" ] }, { "term": "Next-frame prediction", "category": "temporal_world_models", "plain_meaning": "Predicting future visual frames from past frames.", "project_usage": "A field-level world-model objective related to the human-video world-model direction.", "do_not_confuse_with": "Next-action prediction.", "primary_files": [ "docs/data/three_foundation_pipelines.json" ] }, { "term": "Object persistence", "category": "temporal_world_models", "plain_meaning": "Tracking that an object remains present over time even when view or interaction changes.", "project_usage": "Relevant for object-set forecast and long-video reasoning.", "do_not_confuse_with": "A single-frame object detection.", "primary_files": [ "docs/data/task_suite_20.json" ] }, { "term": "Rollout", "category": "temporal_world_models", "plain_meaning": "Repeatedly predicting future steps from a model state.", "project_usage": "Important for judging world models beyond one-step prediction.", "do_not_confuse_with": "A held-out static test row.", "primary_files": [ "docs/data/three_foundation_pipelines.json" ] }, { "term": "Subtask forecasting", "category": "temporal_world_models", "plain_meaning": "Predicting the next higher-level step in an activity.", "project_usage": "Used in the future-task probe line for Qwen3-Omni.", "do_not_confuse_with": "Frame-level action classification.", "primary_files": [ "docs/data/task_method_20_result_matrix.json" ] }, { "term": "Teacher forcing", "category": "temporal_world_models", "plain_meaning": "Training a sequence model using ground-truth previous outputs.", "project_usage": "A likely training option for future sequence/world-model baselines.", "do_not_confuse_with": "Free-running rollout evaluation.", "primary_files": [ "docs/data/foundation_model_plan.json" ] }, { "term": "Temporal leakage", "category": "temporal_world_models", "plain_meaning": "Using future information that would not be available at prediction time.", "project_usage": "Avoided by chronological splits and target-side feature controls.", "do_not_confuse_with": "A low model score.", "primary_files": [ "EVALUATION_PROTOCOL.md" ] }, { "term": "Transition timing", "category": "temporal_world_models", "plain_meaning": "Estimating when the next state or action transition happens.", "project_usage": "Task 20 turns temporal change into a regression target.", "do_not_confuse_with": "Classifying the transition type only.", "primary_files": [ "docs/data/task_suite_20.json" ] }, { "term": "Action chunk", "category": "robotics_vla", "plain_meaning": "A short sequence of low-level actions predicted together.", "project_usage": "The VLA figure and plan use action chunks as the policy-output concept.", "do_not_confuse_with": "A natural-language action label.", "primary_files": [ "docs/data/three_foundation_pipelines.json" ] }, { "term": "Behavior cloning", "category": "robotics_vla", "plain_meaning": "A supervised imitation-learning method for predicting demonstrated actions.", "project_usage": "A plausible baseline once action targets are converted.", "do_not_confuse_with": "Generative video modeling.", "primary_files": [ "docs/data/foundation_model_plan.json" ] }, { "term": "Contact event", "category": "robotics_vla", "plain_meaning": "A moment when a hand, body, or tool touches an object or surface.", "project_usage": "Used in contact-related tasks and action-quality interpretation.", "do_not_confuse_with": "Visual co-occurrence without touch.", "primary_files": [ "docs/data/task_suite_20.json" ] }, { "term": "Dexterity", "category": "robotics_vla", "plain_meaning": "Fine-grained physical manipulation ability.", "project_usage": "Relevant to hand-object interaction, contact, and VLA/policy directions.", "do_not_confuse_with": "High text-generation accuracy.", "primary_files": [ "docs/data/research_directions.json" ] }, { "term": "End effector", "category": "robotics_vla", "plain_meaning": "The robot part that acts on the world, such as a gripper or hand.", "project_usage": "A key target frame for future manipulation-policy conversion.", "do_not_confuse_with": "A camera or global scene coordinate.", "primary_files": [ "docs/data/three_foundation_pipelines.json" ] }, { "term": "Hand-object interaction", "category": "robotics_vla", "plain_meaning": "A physical interaction between hands and objects.", "project_usage": "A central signal family behind action, contact, object relevance, and interaction-text tasks.", "do_not_confuse_with": "Object detection without action.", "primary_files": [ "docs/data/task_suite_20.json" ] }, { "term": "Imitation learning", "category": "robotics_vla", "plain_meaning": "Training a policy to imitate demonstrated behavior.", "project_usage": "Relevant when converting human video/motion into action supervision.", "do_not_confuse_with": "Reinforcement learning from online robot trials.", "primary_files": [ "docs/data/foundation_model_plan.json" ] }, { "term": "Language grounding", "category": "robotics_vla", "plain_meaning": "Connecting text to observed objects, actions, or spatial context.", "project_usage": "Task 8 and VLA directions use language as grounded supervision rather than standalone text.", "do_not_confuse_with": "Caption fluency alone.", "primary_files": [ "docs/data/task_suite_20.json" ] }, { "term": "Policy", "category": "robotics_vla", "plain_meaning": "A mapping from observations to actions.", "project_usage": "A future target for robot-compatible Xperience-derived action data.", "do_not_confuse_with": "A benchmark metric.", "primary_files": [ "docs/data/foundation_model_plan.json" ] }, { "term": "Robot-compatible action target", "category": "robotics_vla", "plain_meaning": "An action representation a robot policy can execute or imitate.", "project_usage": "Needed before OpenVLA/openpi/GR00T-style policy training is meaningful here.", "do_not_confuse_with": "Human-only caption text.", "primary_files": [ "docs/data/foundation_model_plan.json" ] }, { "term": "Vision-language-action model", "category": "robotics_vla", "plain_meaning": "A model that maps visual context and language into actions.", "project_usage": "The VLA direction is a future path after action targets are converted into robot-compatible chunks.", "do_not_confuse_with": "A vision-language model that only answers text.", "primary_files": [ "docs/data/three_foundation_pipelines.json" ] }, { "term": "Compact-proxy score", "category": "tasks_metrics", "plain_meaning": "A bounded proxy metric when a direct raw target is not publicly available.", "project_usage": "Kept explicit in the matrix and gap audit so readers do not over-read it.", "do_not_confuse_with": "A direct target measurement.", "primary_files": [ "TASK_METHOD_20_GAP_AUDIT.md", "docs/data/task_method_20_gap_audit.json" ] }, { "term": "Direct score", "category": "tasks_metrics", "plain_meaning": "A metric computed against the task target directly.", "project_usage": "The preferred score type in the 20-task matrix.", "do_not_confuse_with": "Compact-proxy score.", "primary_files": [ "TASK_METHOD_20_GAP_AUDIT.md", "docs/data/task_method_20_gap_audit.json" ] }, { "term": "Gap audit", "category": "tasks_metrics", "plain_meaning": "A coverage and source-status audit.", "project_usage": "Explains scored, proxy, and unsupported cells.", "do_not_confuse_with": "A performance leaderboard.", "primary_files": [ "TASK_METHOD_20_GAP_AUDIT.md", "docs/data/task_method_20_gap_audit.json" ] }, { "term": "Leakage control", "category": "tasks_metrics", "plain_meaning": "A split or feature rule that prevents using target information unfairly.", "project_usage": "Chronological splits, held-out splits, and source audits protect task interpretation.", "do_not_confuse_with": "Lower training accuracy.", "primary_files": [ "EVALUATION_PROTOCOL.md", "docs/data/evaluation_protocol.json" ] }, { "term": "Normalized radar value", "category": "tasks_metrics", "plain_meaning": "A 0-1 plotting value used only to draw comparable radar polygons.", "project_usage": "Helps visualize metrics with different scales and directions.", "do_not_confuse_with": "The raw metric value to cite.", "primary_files": [ "docs/data/unified_task_model_radar.json", "docs/assets/charts/unified_task_model_radar.svg" ] }, { "term": "Raw metric value", "category": "tasks_metrics", "plain_meaning": "The original metric value emitted by the runner or verified result package.", "project_usage": "This is the value to cite from the 180-result table.", "do_not_confuse_with": "The normalized radar value.", "primary_files": [ "TASK_METHOD_20_RESULT_MATRIX.md", "docs/data/task_method_20_result_matrix.json" ] }, { "term": "Task contract", "category": "tasks_metrics", "plain_meaning": "The definition of one benchmark task.", "project_usage": "Includes input, target/output, metric, split, source artifact, and limitation.", "do_not_confuse_with": "A model architecture.", "primary_files": [ "TASK_SUITE_20.md", "docs/data/task_suite_20.json" ] }, { "term": "Task-method record", "category": "tasks_metrics", "plain_meaning": "One method evaluated on one task.", "project_usage": "9 methods x 20 tasks gives 180 public result records.", "do_not_confuse_with": "A single prediction row.", "primary_files": [ "TASK_METHOD_20_RESULT_MATRIX.md", "docs/data/task_method_20_result_matrix.json" ] }, { "term": "Unified 20-task suite", "category": "tasks_metrics", "plain_meaning": "The current task surface.", "project_usage": "All 20 task contracts are presented together and scored across methods where real artifacts exist.", "do_not_confuse_with": "Historical tier2_task_suite filenames, which are provenance paths rather than a second suite.", "primary_files": [ "TASK_SUITE_20.md", "docs/data/task_suite_20.json" ] }, { "term": "Adapter checkpoint", "category": "training_eval", "plain_meaning": "Saved adapter weights from a fine-tuning run.", "project_usage": "The public model branches publish adapters when validated and public-safe.", "do_not_confuse_with": "Full base-model checkpoint.", "primary_files": [ "docs/data/omni_model_comparison.json" ] }, { "term": "Balanced accuracy", "category": "training_eval", "plain_meaning": "Accuracy averaged across classes to reduce majority-class dominance.", "project_usage": "Useful for imbalanced task labels.", "do_not_confuse_with": "Overall accuracy.", "primary_files": [ "docs/data/task_method_20_result_matrix.json" ] }, { "term": "Chronological split", "category": "training_eval", "plain_meaning": "A split ordered by time.", "project_usage": "Used for the single-episode baselines to reduce future-window leakage.", "do_not_confuse_with": "A random row split.", "primary_files": [ "EVALUATION_PROTOCOL.md" ] }, { "term": "Confusion matrix", "category": "training_eval", "plain_meaning": "A table of predicted classes versus true classes.", "project_usage": "Helps inspect which task labels a method confuses.", "do_not_confuse_with": "A scalar leaderboard score.", "primary_files": [ "results/episode_task_suite/neural_mlp" ] }, { "term": "FSDP", "category": "training_eval", "plain_meaning": "Fully Sharded Data Parallel, a distributed training strategy.", "project_usage": "Appears in full-parameter feasibility and multi-GPU training notes.", "do_not_confuse_with": "A model architecture.", "primary_files": [ "docs/data/qwen3_full_parameter_gates.json" ] }, { "term": "Held-out evaluation", "category": "training_eval", "plain_meaning": "Testing on examples not used for training.", "project_usage": "Required before promoting Qwen/Cosmos results to public evidence.", "do_not_confuse_with": "Training-set loss.", "primary_files": [ "docs/data/omni_model_comparison.json" ] }, { "term": "JSON validity", "category": "training_eval", "plain_meaning": "Whether model output parses as the required JSON schema.", "project_usage": "A key diagnostic for Qwen3-Omni structured-output runs.", "do_not_confuse_with": "Task correctness after parsing.", "primary_files": [ "docs/data/omni_model_comparison.json" ] }, { "term": "Macro F1", "category": "training_eval", "plain_meaning": "The average F1 score across classes, usually treating classes equally.", "project_usage": "Used when class imbalance matters in classification tasks.", "do_not_confuse_with": "Accuracy dominated by frequent classes.", "primary_files": [ "docs/data/task_method_20_result_matrix.json" ] }, { "term": "Mean absolute error", "category": "training_eval", "plain_meaning": "The average absolute difference between predicted and true numeric values.", "project_usage": "Used for regression-style task rows such as timing or trajectory targets.", "do_not_confuse_with": "A classification F1 score.", "primary_files": [ "docs/data/task_method_20_result_matrix.json" ] }, { "term": "Overfit check", "category": "training_eval", "plain_meaning": "A small training test that verifies a model can learn a tiny subset.", "project_usage": "Useful for catching data/model wiring bugs before full training.", "do_not_confuse_with": "Evidence of generalization.", "primary_files": [ "docs/data/foundation_model_plan.json" ] }, { "term": "Parameter-efficient fine-tuning", "category": "training_eval", "plain_meaning": "Updating a small number of added or selected parameters.", "project_usage": "LoRA is the current parameter-efficient path for Qwen/Cosmos branches.", "do_not_confuse_with": "Full-parameter fine-tuning.", "primary_files": [ "docs/data/foundation_model_plan.json" ] }, { "term": "Schema compliance", "category": "training_eval", "plain_meaning": "Whether an output follows the expected field names and value types.", "project_usage": "Needed for structured task probes and public package validation.", "do_not_confuse_with": "High semantic accuracy.", "primary_files": [ "docs/data/omni_model_comparison.json" ] }, { "term": "Smoke run", "category": "training_eval", "plain_meaning": "A short run that checks whether a pipeline can start and execute key steps.", "project_usage": "Used for feasibility gates before expensive full runs.", "do_not_confuse_with": "A complete benchmark result.", "primary_files": [ "docs/data/qwen3_full_parameter_gates.json" ] }, { "term": "Top-k accuracy", "category": "training_eval", "plain_meaning": "A score that counts a prediction correct if the target is among the k highest-ranked outputs.", "project_usage": "Useful for large-label or retrieval-style tasks.", "do_not_confuse_with": "Top-1 exact accuracy.", "primary_files": [ "docs/data/task_method_20_result_matrix.json" ] }, { "term": "Train/validation/test split", "category": "training_eval", "plain_meaning": "A partition that separates model fitting, tuning, and final evaluation examples.", "project_usage": "The selected-128 setup uses a held-out split discipline for model branches.", "do_not_confuse_with": "A random shuffle without temporal or episode boundaries.", "primary_files": [ "EVALUATION_PROTOCOL.md" ] }, { "term": "Cosmos3-Nano", "category": "models_runs", "plain_meaning": "A smaller Cosmos3 compatibility/future-window branch.", "project_usage": "Used for the Nano Future Window row and related diagnostics.", "do_not_confuse_with": "Cosmos3-Super fine-tuned adapter.", "primary_files": [ "docs/data/omni_model_comparison.json" ] }, { "term": "Cosmos3-Super", "category": "models_runs", "plain_meaning": "The larger Cosmos3-style branch tracked in this project.", "project_usage": "Published as Reasoner diagnostics and a separate forward-dynamics LoRA adapter/result branch when verified.", "do_not_confuse_with": "Cosmos3-Nano.", "primary_files": [ "docs/data/omni_model_comparison.json" ] }, { "term": "Foundation pipeline", "category": "models_runs", "plain_meaning": "A high-level training direction.", "project_usage": "Spatial intelligence, human-video world modeling, and vision-language-action are documented as trainable directions with task mappings.", "do_not_confuse_with": "A completed public result row.", "primary_files": [ "THREE_FOUNDATION_PIPELINES.md", "docs/data/three_foundation_pipelines.json" ] }, { "term": "Full-parameter fine-tuning", "category": "models_runs", "plain_meaning": "Updating the whole model rather than only adapters.", "project_usage": "This project records feasibility gates and short pilots, but does not publish full checkpoints.", "do_not_confuse_with": "LoRA adapter publication.", "primary_files": [ "docs/data/qwen3_full_parameter_gates.json" ] }, { "term": "Human-video world model", "category": "models_runs", "plain_meaning": "Learning future frames, actions, and interaction dynamics from human video.", "project_usage": "Uses temporal prediction, next-action, transition, and object-forecast tasks.", "do_not_confuse_with": "Robot policy execution.", "primary_files": [ "THREE_FOUNDATION_PIPELINES.md", "docs/data/three_foundation_pipelines.json" ] }, { "term": "LoRA adapter", "category": "models_runs", "plain_meaning": "A lightweight set of trainable adapter weights.", "project_usage": "Published only when the package is verified and public-safe.", "do_not_confuse_with": "Full base-model weights.", "primary_files": [ "OMNI_MODEL_EXTENSION_CONTRACT.md", "docs/data/omni_model_comparison.json" ] }, { "term": "Metadata baseline", "category": "models_runs", "plain_meaning": "A selected-128 baseline using metadata or text-derived public-safe features.", "project_usage": "Compares simple and neural heads on the held-out split.", "do_not_confuse_with": "Raw video, depth, or audio feature baselines.", "primary_files": [ "docs/data/task_method_20_result_matrix.json" ] }, { "term": "Minimal baseline", "category": "models_runs", "plain_meaning": "A simple non-neural task head; the \"minimum\" reference row in casual wording.", "project_usage": "Provides a reproducible lower-complexity comparison for task feasibility.", "do_not_confuse_with": "Metadata-only selected-128 baseline family.", "primary_files": [ "RESEARCH_TAKEAWAYS.md", "docs/data/task_method_20_result_matrix.json" ] }, { "term": "Neural MLP", "category": "models_runs", "plain_meaning": "A compact neural task head.", "project_usage": "Used for single-episode and selected-128 baseline comparisons.", "do_not_confuse_with": "Foundation-model fine-tuning.", "primary_files": [ "results/episode_task_suite/neural_mlp/", "docs/data/task_method_20_result_matrix.json" ] }, { "term": "Qwen v1-v6", "category": "models_runs", "plain_meaning": "The Qwen3-Omni run lineage.", "project_usage": "v1-v4 are earlier pipeline/ablation evidence, v5 is the prior pinned release, and v6 is the current public 20-task row.", "do_not_confuse_with": "Six different evidence lines.", "primary_files": [ "QWEN3_OMNI_RUN_LINEAGE.md", "docs/data/qwen3_omni_run_lineage.json" ] }, { "term": "Qwen3-Omni", "category": "models_runs", "plain_meaning": "The multimodal foundation-model family used for the Qwen branch.", "project_usage": "The current public 20-task Qwen row is Qwen3-Omni v6 LoRA plus task-specific probes.", "do_not_confuse_with": "Cosmos3 or single-episode task-head baselines.", "primary_files": [ "QWEN3_OMNI_RUN_LINEAGE.md", "docs/data/qwen3_omni_run_lineage.json" ] }, { "term": "Raw-feature baseline", "category": "models_runs", "plain_meaning": "A selected-128 baseline using exported public-safe raw-feature groups.", "project_usage": "Tracks what non-foundation heads can do with richer processed inputs.", "do_not_confuse_with": "Raw gated media redistribution.", "primary_files": [ "docs/data/task_method_20_result_matrix.json" ] }, { "term": "Simple baseline", "category": "models_runs", "plain_meaning": "A non-neural baseline family for the selected-128 rows.", "project_usage": "Used for metadata/text and raw-feature 128-episode comparisons before NN/foundation-model rows.", "do_not_confuse_with": "The single-episode Minimal baseline.", "primary_files": [ "RESEARCH_TAKEAWAYS.md", "docs/data/task_method_20_result_matrix.json" ] }, { "term": "Spatial intelligence", "category": "models_runs", "plain_meaning": "Learning geometry and spatial reasoning from egocentric data.", "project_usage": "Uses video, depth, camera pose, and language tasks to target 3D/space reasoning.", "do_not_confuse_with": "World-model future prediction.", "primary_files": [ "THREE_FOUNDATION_PIPELINES.md", "docs/data/three_foundation_pipelines.json" ] }, { "term": "Vision-language-action", "category": "models_runs", "plain_meaning": "Mapping perception and language to action chunks.", "project_usage": "A future policy/VLA direction that needs action-target conversion and stronger policy packaging.", "do_not_confuse_with": "Qwen3-Omni diagnostic scoring.", "primary_files": [ "THREE_FOUNDATION_PIPELINES.md", "docs/data/three_foundation_pipelines.json" ] }, { "term": "HF artifact dataset", "category": "public_surfaces", "plain_meaning": "Hugging Face dataset repo for derived evidence.", "project_usage": "Stores public-safe reports, metrics, website JSON, and sanitized result packages.", "do_not_confuse_with": "Original Xperience-10M dataset.", "primary_files": [ "ARTIFACT_GUIDE.md", "docs/data/artifact_index.json" ] }, { "term": "HF baseline model repo", "category": "public_surfaces", "plain_meaning": "Hugging Face model repo for lightweight baseline artifacts.", "project_usage": "Mirrors baseline weights, figures, metrics, and task artifacts.", "do_not_confuse_with": "Qwen/Cosmos adapter-specific repos.", "primary_files": [ "PUBLIC_READER_MAP.md", "docs/data/public_reader_map.json" ] }, { "term": "HF Space", "category": "public_surfaces", "plain_meaning": "Hugging Face-hosted app/site surface.", "project_usage": "Mirrors the dashboard and static website assets.", "do_not_confuse_with": "HF artifact dataset or model repo.", "primary_files": [ "PUBLIC_READER_MAP.md", "docs/data/public_reader_map.json" ] }, { "term": "HF weights/results repo", "category": "public_surfaces", "plain_meaning": "A consolidated public-safe model-result bundle.", "project_usage": "Groups baseline weights, verified model artifacts, analysis files, and manifests.", "do_not_confuse_with": "The upstream raw dataset.", "primary_files": [ "PUBLIC_READER_MAP.md" ] }, { "term": "Mirror parity", "category": "public_surfaces", "plain_meaning": "A check that public copies match the source files.", "project_usage": "Records whether GitHub, website, and HF mirrors agree.", "do_not_confuse_with": "A model-quality metric.", "primary_files": [ "docs/data/mirror_parity.json" ] }, { "term": "Public-safe artifact", "category": "public_surfaces", "plain_meaning": "A file that can be mirrored publicly without raw gated content.", "project_usage": "Metrics, JSON summaries, model cards, figures, derived manifests, and approved lightweight weights/adapters.", "do_not_confuse_with": "Raw dataset redistribution.", "primary_files": [ "ARTIFACT_GUIDE.md", "docs/data/artifact_index.json" ] }, { "term": "Publication audit", "category": "public_surfaces", "plain_meaning": "A public-package validation report.", "project_usage": "Confirms required files exist and forbidden raw/private assets are not included.", "do_not_confuse_with": "Scientific peer review.", "primary_files": [ "docs/data/publication_audit.json" ] }, { "term": "Verified package", "category": "public_surfaces", "plain_meaning": "A result or artifact bundle that passed local/public validators.", "project_usage": "Only verified packages are promoted to README, website, and HF surfaces as public evidence.", "do_not_confuse_with": "A running or exploratory experiment.", "primary_files": [ "docs/data/publication_audit.json", "PUBLIC_SURFACE_QA.md" ] } ], "file_entry_points": [ { "need": "Reader navigation", "files": [ "PUBLIC_READER_MAP.md", "docs/data/public_reader_map.json" ] }, { "need": "Task definitions", "files": [ "TASK_SUITE_20.md", "docs/data/task_suite_20.json" ] }, { "need": "Result matrix", "files": [ "TASK_METHOD_20_RESULT_MATRIX.md", "docs/data/task_method_20_result_matrix.json" ] }, { "need": "Direct/proxy status", "files": [ "TASK_METHOD_20_GAP_AUDIT.md", "docs/data/task_method_20_gap_audit.json" ] }, { "need": "Qwen lineage", "files": [ "QWEN3_OMNI_RUN_LINEAGE.md", "docs/data/qwen3_omni_run_lineage.json" ] }, { "need": "128-episode source/features", "files": [ "XPERIENCE10M_128_EPISODE_FEATURE_INDEX.md", "docs/data/xperience10m_128_episode_feature_index.json" ] }, { "need": "Public mirrors", "files": [ "PUBLIC_SURFACE_QA.md", "docs/data/mirror_parity.json", "docs/data/live_publication_status.json" ] } ] }