- 
+ 
Restored direction slide
Vision-language-action models
diff --git a/results/omni_finetune/model_output_task_probes_20260616/RUN_REPORT.md b/results/omni_finetune/model_output_task_probes_20260616/RUN_REPORT.md
index f4e98258c60d5a8cc041b0ea05e9ca1de949dc97..8c123657c0a40efb9a6a1e34d20082f16f705210 100644
--- a/results/omni_finetune/model_output_task_probes_20260616/RUN_REPORT.md
+++ b/results/omni_finetune/model_output_task_probes_20260616/RUN_REPORT.md
@@ -1,6 +1,6 @@
# Existing Model-Output Task Probes
-Generated: `2026-06-18T15:25:22+00:00`
+Generated: `2026-06-18T16:30:30+00:00`
This package scores only task targets already present in verified held-out
prediction JSON. It does not run new inference and does not infer targets that
@@ -10,4 +10,4 @@ are absent from a model branch.
| --- | --- | --- | --- | ---: | ---: | ---: | --- |
| Qwen3-Omni v6 LoRA | qwen3_omni_v6_lora | scored | action_object_relation | n/a | 0.000222 | n/a | results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_multiscale_cap96_v6_rank64_lr5e5_full8gpu_lora_eval_test_full/eval/predictions.jsonl |
| Cosmos3-Super Reasoner | cosmos3_super_reasoner | scored | action_object_relation, time_to_transition | n/a | 0.000000 | 52.946 | results/omni_finetune/verified_public/xperience10m_cosmos3_super_reasoner_128ep_test_full_20260607/eval/predictions.jsonl |
-| Cosmos3-Nano Future Window | cosmos3_nano_future_window | scored | long_horizon_next_action | 0.002491 | n/a | n/a | results/omni_finetune/verified_public/xperience10m_cosmos3_nano_128ep_future_window_h5_compat_adapter_eval_test_full/eval/future_predictions.jsonl |
+| Cosmos3-Nano Future Window | cosmos3_nano_future_window | scored | long_horizon_next_action, modality_reconstruction | 0.002491 | n/a | n/a | results/omni_finetune/verified_public/xperience10m_cosmos3_nano_128ep_future_window_h5_compat_adapter_eval_test_full/eval/future_predictions.jsonl |
diff --git a/results/omni_finetune/model_output_task_probes_20260616/action_object_relation/cosmos3_super_reasoner/metrics.json b/results/omni_finetune/model_output_task_probes_20260616/action_object_relation/cosmos3_super_reasoner/metrics.json
index bcf504c3f6238e9f4f22513c23c6e15d804721c6..99bacbd6b64f1345fd95b40b4c2ed0285203d407 100644
--- a/results/omni_finetune/model_output_task_probes_20260616/action_object_relation/cosmos3_super_reasoner/metrics.json
+++ b/results/omni_finetune/model_output_task_probes_20260616/action_object_relation/cosmos3_super_reasoner/metrics.json
@@ -8,7 +8,7 @@
"predictions_csv": "results/omni_finetune/model_output_task_probes_20260616/action_object_relation/cosmos3_super_reasoner/predictions.csv"
},
"excluded_rows_without_true_relation": 2,
- "generated_at_utc": "2026-06-18T15:25:22+00:00",
+ "generated_at_utc": "2026-06-18T16:30:30+00:00",
"labels": [
"Adjust canned food on shelf :: canned food | cardboard box | store shelf",
"Adjust item on shelf :: shelf | stationery package",
diff --git a/results/omni_finetune/model_output_task_probes_20260616/action_object_relation/qwen3_omni_v6_lora/metrics.json b/results/omni_finetune/model_output_task_probes_20260616/action_object_relation/qwen3_omni_v6_lora/metrics.json
index 77a0a98eb11655786d33bf81a701493d5077059c..c4e06e2ad4d27a6399d8318379e8f2e5f2ffbd8a 100644
--- a/results/omni_finetune/model_output_task_probes_20260616/action_object_relation/qwen3_omni_v6_lora/metrics.json
+++ b/results/omni_finetune/model_output_task_probes_20260616/action_object_relation/qwen3_omni_v6_lora/metrics.json
@@ -8,7 +8,7 @@
"predictions_csv": "results/omni_finetune/model_output_task_probes_20260616/action_object_relation/qwen3_omni_v6_lora/predictions.csv"
},
"excluded_rows_without_true_relation": 18,
- "generated_at_utc": "2026-06-18T15:25:21+00:00",
+ "generated_at_utc": "2026-06-18T16:30:30+00:00",
"labels": [
"Adjust canned food on shelf :: canned food | cardboard box | store shelf",
"Adjust item on shelf :: hand | packaged item | shelf",
diff --git a/results/omni_finetune/model_output_task_probes_20260616/long_horizon_next_action/cosmos3_nano_future_window/metrics.json b/results/omni_finetune/model_output_task_probes_20260616/long_horizon_next_action/cosmos3_nano_future_window/metrics.json
index 4e5f5d94ba03bfaa08d77c2040c3a28b80283578..5e0217efde1aa29c10e018cf6d66cd8b95db0232 100644
--- a/results/omni_finetune/model_output_task_probes_20260616/long_horizon_next_action/cosmos3_nano_future_window/metrics.json
+++ b/results/omni_finetune/model_output_task_probes_20260616/long_horizon_next_action/cosmos3_nano_future_window/metrics.json
@@ -7,7 +7,7 @@
},
"dataset_contract": "xperience10m_future_window_world_model_v0",
"excluded_rows_without_true_action": 0,
- "generated_at_utc": "2026-06-18T15:25:22+00:00",
+ "generated_at_utc": "2026-06-18T16:30:30+00:00",
"horizon_windows": 5,
"known_limitation": "The verified package records horizon_windows rather than a raw wall-clock horizon. This score should be read as the Cosmos-Nano future-window branch for task 13, not as independent proof of an exact five-second raw-video target.",
"labels": [
diff --git a/results/omni_finetune/model_output_task_probes_20260616/modality_reconstruction/cosmos3_nano_future_window/metrics.json b/results/omni_finetune/model_output_task_probes_20260616/modality_reconstruction/cosmos3_nano_future_window/metrics.json
new file mode 100644
index 0000000000000000000000000000000000000000..4f4dc804329384f947c06671c583c46306b6bbdd
--- /dev/null
+++ b/results/omni_finetune/model_output_task_probes_20260616/modality_reconstruction/cosmos3_nano_future_window/metrics.json
@@ -0,0 +1,25 @@
+{
+ "artifact_files": {
+ "metrics_json": "results/omni_finetune/model_output_task_probes_20260616/modality_reconstruction/cosmos3_nano_future_window/metrics.json"
+ },
+ "feature_reconstruction_error": 3479.218317102503,
+ "feature_reconstruction_quality": 0.0002873382957286892,
+ "generated_at_utc": "2026-06-18T16:30:30+00:00",
+ "known_limitation": "The metric is comparable as evidence that the branch emitted a reconstruction objective, but it should not be read as an R2 head trained on the exact simple baseline feature split.",
+ "metric_key": "feature_reconstruction_quality",
+ "model_id": "cosmos3_nano_future_window",
+ "model_label": "Cosmos3-Nano Future Window",
+ "normalization_policy": "This is not the single-episode/128-baseline R2 metric. It is a model-branch reconstruction-quality probe backed by the verified held-out future-window feature reconstruction error.",
+ "num_samples": 378,
+ "primary_metric": "feature_reconstruction_quality",
+ "primary_score": 0.0002873382957286892,
+ "scope": "held_out_test_existing_verified_future_window_reconstruction_metric",
+ "score_policy": "Derived from the verified Cosmos3-Nano future-window package metric. The source package directly reports held-out feature_reconstruction_error; this artifact maps it onto task 10 as an inverse reconstruction-quality score 1 / (1 + error) so the matrix can retain its higher-is-better convention.",
+ "source_metric_key": "feature_reconstruction_error",
+ "source_metrics_json": "results/omni_finetune/verified_public/xperience10m_cosmos3_nano_128ep_future_window_h5_compat_adapter_eval_test_full/eval/metrics.json",
+ "status": "pass",
+ "task_id": "modality_reconstruction",
+ "task_label": "Cross-Modal Reconstruction",
+ "task_number": 10,
+ "title": "Cosmos3-Nano Future Window Modality Reconstruction Probe"
+}
diff --git a/results/omni_finetune/model_output_task_probes_20260616/summary.json b/results/omni_finetune/model_output_task_probes_20260616/summary.json
index 1a12271438072214db8f7431ed7d9be3cb6cc7b6..73856e1caf1ce5eba084ffeb8b8a402ab706fbd2 100644
--- a/results/omni_finetune/model_output_task_probes_20260616/summary.json
+++ b/results/omni_finetune/model_output_task_probes_20260616/summary.json
@@ -1,5 +1,5 @@
{
- "generated_at_utc": "2026-06-18T15:25:22+00:00",
+ "generated_at_utc": "2026-06-18T16:30:30+00:00",
"methods": {
"cosmos3_nano_future_window": {
"label": "Cosmos3-Nano Future Window",
@@ -13,6 +13,13 @@
"long_horizon_next_action_macro_f1": 0.0024906600249066007,
"scored_rows": 378,
"source_metrics_json": "results/omni_finetune/model_output_task_probes_20260616/long_horizon_next_action/cosmos3_nano_future_window/metrics.json"
+ },
+ "modality_reconstruction": {
+ "feature_reconstruction_error": 3479.218317102503,
+ "feature_reconstruction_quality": 0.0002873382957286892,
+ "num_samples": 378,
+ "source_metrics_json": "results/omni_finetune/model_output_task_probes_20260616/modality_reconstruction/cosmos3_nano_future_window/metrics.json",
+ "source_verified_metrics_json": "results/omni_finetune/verified_public/xperience10m_cosmos3_nano_128ep_future_window_h5_compat_adapter_eval_test_full/eval/metrics.json"
}
},
"unsupported_tasks": {
@@ -59,11 +66,12 @@
}
},
"scope": "Task-specific scoring from existing verified held-out model outputs. No new model inference, training, or target backfilling is performed.",
- "scored_method_task_count_added": 4,
+ "scored_method_task_count_added": 5,
"status": "pass",
"task_ids_added_to_matrix": [
"action_object_relation",
"long_horizon_next_action",
+ "modality_reconstruction",
"time_to_transition"
],
"title": "Existing Model-Output Task Probes"
diff --git a/results/omni_finetune/model_output_task_probes_20260616/time_to_transition/cosmos3_super_reasoner/metrics.json b/results/omni_finetune/model_output_task_probes_20260616/time_to_transition/cosmos3_super_reasoner/metrics.json
index 7a779d693863bc75b2f1d1c60839c8a36ccd2d73..87c1006a0abbefc572012687411c01699161ddf2 100644
--- a/results/omni_finetune/model_output_task_probes_20260616/time_to_transition/cosmos3_super_reasoner/metrics.json
+++ b/results/omni_finetune/model_output_task_probes_20260616/time_to_transition/cosmos3_super_reasoner/metrics.json
@@ -5,7 +5,7 @@
},
"cap_frames": 200,
"excluded_rows_without_true_action": 0,
- "generated_at_utc": "2026-06-18T15:25:22+00:00",
+ "generated_at_utc": "2026-06-18T16:30:30+00:00",
"known_limitation": "This is a derived action-sequence probe, not evidence of a separately trained time-regression head. It is included because task 20's target is deterministically derivable from a sequence of action labels.",
"metric_direction": "lower",
"metric_key": "time_to_transition_mae",
diff --git a/scripts/build_unified_task_model_radar.py b/scripts/build_unified_task_model_radar.py
index 7355a8ecbb8861045c4605c661fe59a37fb5ceb9..d544f5a7934f6421302011e9f9dad14202b11e71 100644
--- a/scripts/build_unified_task_model_radar.py
+++ b/scripts/build_unified_task_model_radar.py
@@ -60,6 +60,11 @@ QWEN_CROSS_MODAL_RETRIEVAL_PROBE_DIR = (
/ "results/omni_finetune"
/ "xperience10m_qwen3_omni_v6_cross_modal_retrieval_probe_a100_20260618T000000Z"
)
+QWEN_CAMERA_VIEW_SYNC_PROBE_DIR = (
+ ROOT
+ / "results/omni_finetune"
+ / "xperience10m_qwen3_omni_v6_camera_view_sync_probe_a100_20260619T000000Z"
+)
QWEN_ACTION_OBJECT_METRICS_PATH = (
MODEL_OUTPUT_TASK_PROBE_DIR / "action_object_relation/qwen3_omni_v6_lora/metrics.json"
)
@@ -72,6 +77,9 @@ COSMOS_SUPER_TIME_TO_TRANSITION_METRICS_PATH = (
COSMOS_NANO_LONG_HORIZON_METRICS_PATH = (
MODEL_OUTPUT_TASK_PROBE_DIR / "long_horizon_next_action/cosmos3_nano_future_window/metrics.json"
)
+COSMOS_NANO_MODALITY_RECONSTRUCTION_METRICS_PATH = (
+ MODEL_OUTPUT_TASK_PROBE_DIR / "modality_reconstruction/cosmos3_nano_future_window/metrics.json"
+)
QWEN_FUTURE_TASK_METRIC_PATHS = {
"caption_grounding": QWEN_RETRIEVAL_TASK_PROBE_DIR / "caption_grounding/metrics.json",
"cross_modal_retrieval": QWEN_CROSS_MODAL_RETRIEVAL_PROBE_DIR / "cross_modal_retrieval/metrics.json",
@@ -81,6 +89,7 @@ QWEN_FUTURE_TASK_METRIC_PATHS = {
"next_subtask_forecast": QWEN_FUTURE_TASK_PROBE_DIR / "next_subtask_forecast/metrics.json",
"object_set_forecast": QWEN_FUTURE_TASK_PROBE_DIR / "object_set_forecast/metrics.json",
"time_to_transition": QWEN_ORDER_SYNC_TIME_PROBE_DIR / "time_to_transition/metrics.json",
+ "camera_view_sync_retrieval": QWEN_CAMERA_VIEW_SYNC_PROBE_DIR / "camera_view_sync_retrieval/metrics.json",
}
QWEN_FUTURE_TASK_METRIC_KEYS = {
"caption_grounding": "caption_grounding_mrr",
@@ -91,6 +100,7 @@ QWEN_FUTURE_TASK_METRIC_KEYS = {
"next_subtask_forecast": "next_subtask_forecast_macro_f1",
"object_set_forecast": "object_set_forecast_micro_f1",
"time_to_transition": "time_to_transition_mae",
+ "camera_view_sync_retrieval": "camera_view_sync_retrieval_mrr",
}
OUTPUT_JSON = ROOT / "docs/data/unified_task_model_radar.json"
OUTPUT_SINGLE_JSON = ROOT / "docs/data/single_episode_task_model_radar.json"
@@ -213,6 +223,9 @@ FOUNDATION_TASK_METRICS = {
"long_horizon_next_action": {
"cosmos3_nano_future_window": "long_horizon_next_action_macro_f1",
},
+ "modality_reconstruction": {
+ "cosmos3_nano_future_window": "feature_reconstruction_quality",
+ },
"cross_modal_retrieval": {
"cosmos3_nano_future_window": "future_retrieval_mrr",
},
@@ -238,7 +251,9 @@ FOUNDATION_METRIC_SOURCE_OVERRIDES = {
("qwen3_omni_v6_lora", "next_subtask_forecast"): QWEN_FUTURE_TASK_METRIC_PATHS["next_subtask_forecast"],
("qwen3_omni_v6_lora", "object_set_forecast"): QWEN_FUTURE_TASK_METRIC_PATHS["object_set_forecast"],
("qwen3_omni_v6_lora", "time_to_transition"): QWEN_FUTURE_TASK_METRIC_PATHS["time_to_transition"],
+ ("qwen3_omni_v6_lora", "camera_view_sync_retrieval"): QWEN_FUTURE_TASK_METRIC_PATHS["camera_view_sync_retrieval"],
("cosmos3_nano_future_window", "long_horizon_next_action"): COSMOS_NANO_LONG_HORIZON_METRICS_PATH,
+ ("cosmos3_nano_future_window", "modality_reconstruction"): COSMOS_NANO_MODALITY_RECONSTRUCTION_METRICS_PATH,
("cosmos3_super_reasoner", "time_to_transition"): COSMOS_SUPER_TIME_TO_TRANSITION_METRICS_PATH,
}
@@ -274,7 +289,7 @@ METHOD_DETAILS = {
"raw128_neural_mlp": "128-episode 4430-dim sensor NPZ MLP heads; tasks 15/19 use compact proxies.",
"qwen3_omni_v6_lora": "Verified held-out Qwen3-Omni v6 LoRA metrics, plus task 16 and any completed private-GPU future-task probes scored from task-specific JSON.",
"cosmos3_super_reasoner": "Verified Cosmos3-Super base-weight Reasoner JSON-task evaluation, plus task 16 and a derived task-20 action-boundary timing probe scored from existing verified JSON.",
- "cosmos3_nano_future_window": "Verified Cosmos3-Nano future-window compatibility metrics, plus task 13 scored from existing held-out future-action predictions.",
+ "cosmos3_nano_future_window": "Verified Cosmos3-Nano future-window compatibility metrics, plus task 10 reconstruction quality and task 13 scored from existing held-out future-window artifacts.",
}
PROXY_TASK_IDS = {"interaction_text_prediction", "camera_view_sync_retrieval"}
@@ -632,6 +647,7 @@ def build_payload() -> dict[str, Any]:
cosmos_super.update(read_json(COSMOS_SUPER_ACTION_OBJECT_METRICS_PATH))
cosmos_super.update(read_json(COSMOS_SUPER_TIME_TO_TRANSITION_METRICS_PATH))
cosmos_nano.update(read_json(COSMOS_NANO_LONG_HORIZON_METRICS_PATH))
+ cosmos_nano.update(read_json(COSMOS_NANO_MODALITY_RECONSTRUCTION_METRICS_PATH))
foundation_task_metrics = foundation_task_metric_mapping(qwen)
foundation_metrics = {
"qwen3_omni_v6_lora": qwen,
diff --git a/scripts/omni/eval_qwen3_omni_retrieval_task_probes.py b/scripts/omni/eval_qwen3_omni_retrieval_task_probes.py
index 22c8cb8088e026a061f82ed192f19b4da2c90ff6..71654b3e13bf9ec938af8dfbe0e8ad9c28ae12b3 100644
--- a/scripts/omni/eval_qwen3_omni_retrieval_task_probes.py
+++ b/scripts/omni/eval_qwen3_omni_retrieval_task_probes.py
@@ -15,6 +15,8 @@ import csv
import hashlib
import json
import re
+import shutil
+import subprocess
import time
from collections import OrderedDict
from pathlib import Path
@@ -49,6 +51,16 @@ TASK_SPECS: OrderedDict[str, dict[str, Any]] = OrderedDict(
"prediction_key": "ranked_candidates",
},
),
+ (
+ "camera_view_sync_retrieval",
+ {
+ "task_number": 19,
+ "label": "Camera-View Sync Retrieval",
+ "family": "retrieval",
+ "metric_key": "camera_view_sync_retrieval_mrr",
+ "prediction_key": "ranked_candidates",
+ },
+ ),
]
)
@@ -161,6 +173,26 @@ def media_video_path(sample: dict[str, Any]) -> str | None:
return media.get("mosaic_video_path") or sample.get("primary_video_path")
+def camera_view_paths(sample: dict[str, Any]) -> list[dict[str, str]]:
+ media = sample.get("media") if isinstance(sample.get("media"), dict) else {}
+ values = media.get("video_paths")
+ if not isinstance(values, list):
+ return []
+ views: list[dict[str, str]] = []
+ for item in values:
+ if not isinstance(item, dict):
+ continue
+ name = normalize_text(item.get("name"))
+ path = normalize_text(item.get("path"))
+ if name and path:
+ views.append({"name": name, "path": path})
+ return views
+
+
+def has_camera_view_pair(sample: dict[str, Any]) -> bool:
+ return len(camera_view_paths(sample)) >= 2
+
+
def parse_json_object(text: str) -> dict[str, Any]:
raw = str(text or "").strip()
if raw.startswith("```"):
@@ -226,6 +258,21 @@ def build_candidate_indices(
if candidate_count < 2 or candidate_count > 8:
raise ValueError("--candidate-count must be between 2 and 8")
sample = samples[sample_idx]
+ if task_id == "camera_view_sync_retrieval":
+ negatives = [
+ idx
+ for idx in eval_pool
+ if idx != sample_idx
+ and has_camera_view_pair(samples[idx])
+ and (
+ samples[idx].get("episode_id") != sample.get("episode_id")
+ or row_start(samples[idx]) != row_start(sample)
+ )
+ ]
+ negatives.sort(key=lambda idx: stable_score(task_id, sample.get("id"), samples[idx].get("id")))
+ selected = [sample_idx] + negatives[: candidate_count - 1]
+ selected.sort(key=lambda idx: stable_score(task_id, "order", sample.get("id"), samples[idx].get("id")))
+ return selected
true_action = normalize_text(answer(sample).get("action")).casefold()
true_episode = sample.get("episode_id")
negatives = [
@@ -244,6 +291,91 @@ def build_candidate_indices(
return selected
+def reference_camera_view(sample: dict[str, Any]) -> dict[str, str]:
+ views = camera_view_paths(sample)
+ if len(views) < 2:
+ raise ValueError(f"sample lacks paired camera views: {sample.get('id')}")
+ return views[0]
+
+
+def candidate_camera_view(sample: dict[str, Any]) -> dict[str, str]:
+ views = camera_view_paths(sample)
+ if len(views) < 2:
+ raise ValueError(f"sample lacks paired camera views: {sample.get('id')}")
+ return views[1]
+
+
+def camera_view_clip_path(sample: dict[str, Any], view: dict[str, str], clip_dir: Path) -> str:
+ start = row_start(sample)
+ end = row_end(sample)
+ source = Path(view["path"])
+ if end < start:
+ raise ValueError(f"invalid frame window for {sample.get('id')}: {start}-{end}")
+ digest = hashlib.sha1(
+ f"{sample.get('id')}::{view['name']}::{source}::{start}:{end}".encode("utf-8")
+ ).hexdigest()[:16]
+ safe_view = re.sub(r"[^A-Za-z0-9_.-]+", "_", view["name"]).strip("_") or "camera"
+ output = clip_dir / f"{digest}_{safe_view}_{start}_{end}.mp4"
+ if output.exists() and output.stat().st_size > 0:
+ return str(output)
+ if not source.exists():
+ raise FileNotFoundError(f"camera source video not found: {source}")
+ output.parent.mkdir(parents=True, exist_ok=True)
+ if shutil.which("ffmpeg"):
+ frame_filter = f"select=between(n\\,{start}\\,{end}),setpts=N/FRAME_RATE/TB"
+ subprocess.run(
+ [
+ "ffmpeg",
+ "-hide_banner",
+ "-loglevel",
+ "error",
+ "-y",
+ "-i",
+ str(source),
+ "-vf",
+ frame_filter,
+ "-an",
+ "-pix_fmt",
+ "yuv420p",
+ str(output),
+ ],
+ check=True,
+ )
+ else:
+ import cv2
+
+ cap = cv2.VideoCapture(str(source))
+ if not cap.isOpened():
+ raise RuntimeError(f"unable to open camera source video: {source}")
+ fps = cap.get(cv2.CAP_PROP_FPS) or 30.0
+ width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH) or 0)
+ height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT) or 0)
+ if width <= 0 or height <= 0:
+ cap.release()
+ raise RuntimeError(f"invalid camera source dimensions for {source}: {width}x{height}")
+ writer = cv2.VideoWriter(str(output), cv2.VideoWriter_fourcc(*"mp4v"), float(fps), (width, height))
+ if not writer.isOpened():
+ cap.release()
+ raise RuntimeError(f"unable to open camera clip writer: {output}")
+ cap.set(cv2.CAP_PROP_POS_FRAMES, start)
+ frame_index = start
+ written = 0
+ while frame_index <= end:
+ ok, frame = cap.read()
+ if not ok:
+ break
+ writer.write(frame)
+ written += 1
+ frame_index += 1
+ writer.release()
+ cap.release()
+ if written == 0:
+ raise RuntimeError(f"no frames written for camera clip: {source} frames {start}-{end}")
+ if not output.exists() or output.stat().st_size == 0:
+ raise RuntimeError(f"failed to build camera clip: {output}")
+ return str(output)
+
+
def query_text(sample: dict[str, Any]) -> str:
payload = answer(sample)
objects = payload.get("objects") if isinstance(payload.get("objects"), list) else []
@@ -311,6 +443,23 @@ def sensor_query_text(sample: dict[str, Any], cache: SensorFeatureCache) -> str:
return "\n".join(lines)
+def artifact_query_text(task_id: str, sample: dict[str, Any], sensor_cache: SensorFeatureCache | None) -> str:
+ if task_id == "cross_modal_retrieval":
+ if sensor_cache is None:
+ raise ValueError("cross_modal_retrieval requires a sensor feature cache")
+ return sensor_query_text(sample, sensor_cache)
+ if task_id == "camera_view_sync_retrieval":
+ ref = reference_camera_view(sample)
+ return "\n".join(
+ [
+ f"Reference camera view: {ref['name']}",
+ f"Window frames: {row_start(sample)}-{row_end(sample)}",
+ "Target is a different camera view from the same synchronized window.",
+ ]
+ )
+ return query_text(sample)
+
+
def build_messages(
samples: list[dict[str, Any]],
sample_idx: int,
@@ -318,6 +467,7 @@ def build_messages(
task_id: str,
spec: dict[str, Any],
sensor_cache: SensorFeatureCache | None = None,
+ camera_clip_dir: Path | None = None,
) -> tuple[list[dict[str, Any]], str, list[dict[str, Any]]]:
letters = [chr(ord("A") + pos) for pos in range(len(candidate_indices))]
true_letter = letters[candidate_indices.index(sample_idx)]
@@ -328,6 +478,20 @@ def build_messages(
task_instruction = "Rank the candidate video windows by which one is synchronized with the sensor/motion query."
query = sensor_query_text(samples[sample_idx], sensor_cache)
query_header = "Sensor/motion query:"
+ elif task_id == "camera_view_sync_retrieval":
+ task_instruction = (
+ "Rank the candidate camera-view clips by which one is synchronized with the reference clip. "
+ "The correct candidate shows the same time window from a different camera; distractors are other windows."
+ )
+ ref = reference_camera_view(samples[sample_idx])
+ query = "\n".join(
+ [
+ f"Reference camera view: {ref['name']}",
+ f"Window frames: {row_start(samples[sample_idx])}-{row_end(samples[sample_idx])}",
+ "Use visual timing, hands, objects, and scene motion. Do not use action, subtask, or object labels.",
+ ]
+ )
+ query_header = "Reference clip:"
else:
task_instruction = "Rank the candidate video windows by how well they match the text query."
query = query_text(samples[sample_idx])
@@ -349,8 +513,22 @@ def build_messages(
),
}
]
+ if task_id == "camera_view_sync_retrieval":
+ if camera_clip_dir is None:
+ raise ValueError("camera_view_sync_retrieval requires a camera clip directory")
+ ref_view = reference_camera_view(samples[sample_idx])
+ content.append({"type": "video", "video": camera_view_clip_path(samples[sample_idx], ref_view, camera_clip_dir)})
for letter, idx in zip(letters, candidate_indices):
sample = samples[idx]
+ if task_id == "camera_view_sync_retrieval":
+ if camera_clip_dir is None:
+ raise ValueError("camera_view_sync_retrieval requires a camera clip directory")
+ view = candidate_camera_view(sample)
+ candidate_video = camera_view_clip_path(sample, view, camera_clip_dir)
+ candidate_view_name = view["name"]
+ else:
+ candidate_video = media_video_path(sample)
+ candidate_view_name = "mosaic"
candidate_records.append(
{
"letter": letter,
@@ -358,11 +536,12 @@ def build_messages(
"episode_id": sample.get("episode_id"),
"start_frame": row_start(sample),
"end_frame": row_end(sample),
+ "view_name": candidate_view_name,
"is_target": idx == sample_idx,
}
)
content.append({"type": "text", "text": f"Candidate {letter} video window:"})
- content.append({"type": "video", "video": media_video_path(sample)})
+ content.append({"type": "video", "video": candidate_video})
return (
[
{"role": "system", "content": [{"type": "text", "text": SYSTEM_PROMPT}]},
@@ -447,6 +626,7 @@ def score_retrieval(rows: list[dict[str, Any]]) -> dict[str, float]:
"mrr": mrr,
"caption_grounding_mrr": mrr,
"cross_modal_retrieval_mrr": mrr,
+ "camera_view_sync_retrieval_mrr": mrr,
"top1_accuracy": top1 / len(rows) if rows else 0.0,
}
@@ -494,6 +674,13 @@ def score_task(task_id: str, spec: dict[str, Any], rows: list[dict[str, Any]], o
"candidates are shuffled staged mosaic video windows, and the score is MRR of the "
"synchronized true window. No action/subtask/object labels are included in the query."
)
+ elif task_id == "camera_view_sync_retrieval":
+ score_policy = (
+ "GPU-backed Qwen3-Omni v6 camera-view synchronization retrieval probe. The prompt shows "
+ "one raw camera view as the reference and asks the model to rank shuffled raw candidate "
+ "views; the true target is a different camera from the same held-out time window. No "
+ "action, subtask, object, or future labels are included."
+ )
else:
score_policy = (
"GPU-backed Qwen3-Omni v6 text-to-video retrieval probe. The text query is built "
@@ -540,6 +727,9 @@ def main() -> int:
if "cross_modal_retrieval" in selected_tasks:
eval_indices = [idx for idx in eval_indices if has_sensor_feature(samples[idx])]
eval_pool = [idx for idx in eval_pool if has_sensor_feature(samples[idx])]
+ if "camera_view_sync_retrieval" in selected_tasks:
+ eval_indices = [idx for idx in eval_indices if has_camera_view_pair(samples[idx])]
+ eval_pool = [idx for idx in eval_pool if has_camera_view_pair(samples[idx])]
if not eval_indices:
raise ValueError("No evaluation samples with retrieval candidates selected.")
@@ -559,6 +749,7 @@ def main() -> int:
model, processor = load_model_processor(args)
sensor_cache = SensorFeatureCache() if "cross_modal_retrieval" in selected_tasks else None
+ camera_clip_dir = args.output_dir / "camera_view_sync_clips" if "camera_view_sync_retrieval" in selected_tasks else None
partial_by_task = {
task_id: {
row.get("prediction_id"): row
@@ -585,6 +776,7 @@ def main() -> int:
task_id,
spec,
sensor_cache=sensor_cache,
+ camera_clip_dir=camera_clip_dir,
)
raw = generate_messages(model, processor, messages, args)
valid_letters = [record["letter"] for record in candidate_records]
@@ -599,7 +791,7 @@ def main() -> int:
"episode_id": sample.get("episode_id"),
"start_frame": row_start(sample),
"end_frame": row_end(sample),
- "query_text": sensor_query_text(sample, sensor_cache) if task_id == "cross_modal_retrieval" else query_text(sample),
+ "query_text": artifact_query_text(task_id, sample, sensor_cache),
"candidates": candidate_records,
"true_letter": true_letter,
"predicted_ranking": ranking,
diff --git a/scripts/omni/score_existing_model_output_task_probes.py b/scripts/omni/score_existing_model_output_task_probes.py
index 97633df607e6bddb45f4713c3d9e71d79e34d06b..31f3cb2c2037c349ab73d6e7338996e865a80a59 100644
--- a/scripts/omni/score_existing_model_output_task_probes.py
+++ b/scripts/omni/score_existing_model_output_task_probes.py
@@ -48,6 +48,11 @@ MODEL_SPECS = {
"xperience10m_cosmos3_nano_128ep_future_window_h5_compat_adapter_eval_test_full/"
"eval/future_predictions.jsonl"
),
+ "metrics_json": (
+ "results/omni_finetune/verified_public/"
+ "xperience10m_cosmos3_nano_128ep_future_window_h5_compat_adapter_eval_test_full/"
+ "eval/metrics.json"
+ ),
"dataset_manifest": (
"results/omni_finetune/verified_public/"
"xperience10m_cosmos3_nano_128ep_future_window_h5_compat_adapter_eval_test_full/"
@@ -386,6 +391,61 @@ def score_cosmos_nano_long_horizon_next_action(
return metrics
+def score_modality_reconstruction_from_feature_error(
+ *,
+ model_id: str,
+ spec: dict[str, Any],
+ metrics_json: Path,
+ output_dir: Path,
+ workspace: Path,
+) -> dict[str, Any]:
+ source_metrics = json.loads(metrics_json.read_text(encoding="utf-8"))
+ error = source_metrics.get("feature_reconstruction_error")
+ if not isinstance(error, (int, float)):
+ raise RuntimeError(f"feature_reconstruction_error is absent from {metrics_json}")
+ quality = 1.0 / (1.0 + float(error)) if error >= 0 else 0.0
+ metrics = {
+ "title": f"{spec['label']} Modality Reconstruction Probe",
+ "status": "pass",
+ "generated_at_utc": datetime.now(timezone.utc).isoformat(timespec="seconds"),
+ "model_id": model_id,
+ "model_label": spec["label"],
+ "task_id": "modality_reconstruction",
+ "task_number": 10,
+ "task_label": "Cross-Modal Reconstruction",
+ "metric_key": "feature_reconstruction_quality",
+ "primary_metric": "feature_reconstruction_quality",
+ "primary_score": quality,
+ "feature_reconstruction_quality": quality,
+ "feature_reconstruction_error": float(error),
+ "source_metric_key": "feature_reconstruction_error",
+ "source_metrics_json": relpath(metrics_json, workspace),
+ "scope": "held_out_test_existing_verified_future_window_reconstruction_metric",
+ "score_policy": (
+ "Derived from the verified Cosmos3-Nano future-window package metric. The "
+ "source package directly reports held-out feature_reconstruction_error; this "
+ "artifact maps it onto task 10 as an inverse reconstruction-quality score "
+ "1 / (1 + error) so the matrix can retain its higher-is-better convention."
+ ),
+ "normalization_policy": (
+ "This is not the single-episode/128-baseline R2 metric. It is a model-branch "
+ "reconstruction-quality probe backed by the verified held-out future-window "
+ "feature reconstruction error."
+ ),
+ "known_limitation": (
+ "The metric is comparable as evidence that the branch emitted a reconstruction "
+ "objective, but it should not be read as an R2 head trained on the exact simple "
+ "baseline feature split."
+ ),
+ "num_samples": source_metrics.get("num_samples"),
+ "artifact_files": {
+ "metrics_json": relpath(output_dir / "metrics.json", workspace),
+ },
+ }
+ write_json(output_dir / "metrics.json", metrics)
+ return metrics
+
+
def row_start(row: dict[str, Any]) -> int:
window = row.get("center_window") if isinstance(row.get("center_window"), dict) else {}
return int(window.get("start_frame", 0) or 0)
@@ -644,6 +704,21 @@ def main() -> int:
"long_horizon_next_action_macro_f1": metrics["long_horizon_next_action_macro_f1"],
"long_horizon_next_action_accuracy": metrics["long_horizon_next_action_accuracy"],
}
+ metrics_path = workspace / spec["metrics_json"]
+ metrics = score_modality_reconstruction_from_feature_error(
+ model_id=model_id,
+ spec=spec,
+ metrics_json=metrics_path,
+ output_dir=output_dir / "modality_reconstruction" / model_id,
+ workspace=workspace,
+ )
+ task_results["modality_reconstruction"] = {
+ "source_metrics_json": metrics["artifact_files"]["metrics_json"],
+ "source_verified_metrics_json": metrics["source_metrics_json"],
+ "feature_reconstruction_quality": metrics["feature_reconstruction_quality"],
+ "feature_reconstruction_error": metrics["feature_reconstruction_error"],
+ "num_samples": metrics.get("num_samples"),
+ }
if spec.get("time_to_transition_from_action_sequence"):
metrics = score_time_to_transition_from_action_sequence(
model_id=model_id,
@@ -676,7 +751,12 @@ def main() -> int:
"Task-specific scoring from existing verified held-out model outputs. "
"No new model inference, training, or target backfilling is performed."
),
- "task_ids_added_to_matrix": ["action_object_relation", "long_horizon_next_action", "time_to_transition"],
+ "task_ids_added_to_matrix": [
+ "action_object_relation",
+ "long_horizon_next_action",
+ "modality_reconstruction",
+ "time_to_transition",
+ ],
"scored_method_task_count_added": scored_count,
"methods": methods,
}
+ 
+ 