Deliver verified Stage-3 Small: 1M train, 50K validation, 50K test

README.md

@@ -37,7 +37,7 @@ configs:
 <!-- STAGE3_SMALL_START -->
 ## Small: prior-restoration experiment
 
-**Ready:** 1,000,000 train + 50,000 validation + 50,000 test rows. Use config `small`, pinned tag `small-v1.0`. [Design, exact task/source/modality counts and loading guide](small/README.md). Uses normalized answer targets and a grounding-aware media loader. **The original media_complete config includes these held-out groups; do not train it unfiltered when evaluating on Small.**
+**Ready:** 1,000,000 train + 50,000 validation + 50,000 test rows. Use config `small`, pinned tag `small-v1.0.1`. [Design, exact task/source/modality counts and loading guide](small/README.md). Uses normalized answer targets and a grounding-aware media loader. **The original media_complete config includes these held-out groups; do not train it unfiltered when evaluating on Small.**
 
 <!-- STAGE3_SMALL_END -->
 

small/README.md

@@ -86,7 +86,7 @@ Small 是独立 HF config `small`。原 `media_complete` 保持不变，**它包
 
 ```python
 from datasets import load_dataset
-ds = load_dataset('AnchorSR/TrainingData_Stage3', 'small', revision='small-v1.0')
+ds = load_dataset('AnchorSR/TrainingData_Stage3', 'small', revision='small-v1.0.1')
 ```
 
 实际读取图像/视频，请下载 `small/tools/load_stage3_small.py` 和同目录 `small_training.py`，放在同一目录。依赖：`huggingface_hub pyarrow requests pillow numpy`；视频解码另需 `av` 或你的模型视频处理器。
@@ -94,7 +94,7 @@ ds = load_dataset('AnchorSR/TrainingData_Stage3', 'small', revision='small-v1.0'
 ```python
 from load_stage3_small import SmallDataset
 data = SmallDataset('/scratch/e1503372/small_training_cache',
-                    revision='small-v1.0', max_cache_gib=80)
+                    revision='small-v1.0.1', max_cache_gib=80)
 row = next(data.rows('train'))
 sample = data.sample(row)
 # sample: question, answer, images (PIL, 源帧顺序), videos (本地文件路径)

small/code-tests.json

@@ -1,7 +1,7 @@
 {
   "scope": "Stage-3 tests",
   "exit_code": 0,
-  "output": ".....................................                                    [100%]\n37 passed in 1.51s\n",
+  "output": ".....................................                                    [100%]\n37 passed in 3.66s\n",
   "stderr": "",
   "note": "Unrelated ScanNet++ full-repository tests require cv2 and a newer Python; no claim of whole-repository test success."
 }

small/tools/build_stage3_small.py

@@ -26,6 +26,12 @@ FAMILY_TARGETS = {"object_distance": 510000, "object_dimension": 300000,
                   "camera_pose_measurement": 15000, "room_measurement": 5000}
 
 
+def small_parquets():
+    # Do not include manifests/group-splits.parquet when rerunning verification.
+    return '['+','.join("'"+str(ROOT/f'release/small/{split}/*.parquet')+"'"
+                        for split in ('train','validation','test'))+']'
+
+
 def save(path, obj):
     path.parent.mkdir(parents=True, exist_ok=True)
     tmp = path.with_suffix(path.suffix + ".tmp")
@@ -439,7 +445,7 @@ def index_assets():
         f.rfilename:f.size for f in HfApi().dataset_info(REPO,revision=json.loads((ROOT/'input_manifest.json').read_text())['revision'],files_metadata=True).siblings
         if f.rfilename.startswith('media/')}
     conn=db()
-    conn.execute(f"CREATE OR REPLACE VIEW small AS SELECT * FROM read_parquet('{ROOT}/release/small/*/*.parquet')")
+    conn.execute(f"CREATE OR REPLACE VIEW small AS SELECT * FROM read_parquet({small_parquets()})")
     conn.execute('CREATE OR REPLACE TABLE needed AS SELECT DISTINCT source,unnest(media).path media_path FROM small')
     conn.execute(f"""COPY (SELECT a.* FROM read_parquet('{ROOT}/all_asset_offsets.parquet') a
           SEMI JOIN needed USING(source,media_path)
@@ -463,7 +469,7 @@ def verify():
     import pyarrow.parquet as pq
     conn=db()
     output=ROOT/'release'
-    conn.execute(f"CREATE OR REPLACE VIEW small AS SELECT * FROM read_parquet('{output}/small/*/*.parquet')")
+    conn.execute(f"CREATE OR REPLACE VIEW small AS SELECT * FROM read_parquet({small_parquets()})")
     checks={
         'duplicate_sample_ids':"SELECT count(*)-count(DISTINCT sample_id) FROM small",
         'group_split_overlap':"SELECT count(*) FROM (SELECT group_id FROM small GROUP BY 1 HAVING count(DISTINCT split)>1)",
@@ -488,7 +494,7 @@ def verify():
     results['seed']=SEED
     import math
     grounding_rows=0
-    for path in sorted((output/'small').glob('*/*.parquet')):
+    for path in sorted(p for split in ('train','validation','test') for p in (output/'small'/split).glob('*.parquet')):
         for batch in pq.ParquetFile(path).iter_batches(columns=['sample_id','media','spatial_annotations'],batch_size=1000):
             for row in batch.to_pylist():
                 if not row['spatial_annotations']:continue
@@ -528,7 +534,7 @@ def smoke():
     from load_stage3_small import SmallDataset
     conn=db()
     output=ROOT/'release'
-    conn.execute(f"CREATE OR REPLACE VIEW small AS SELECT * FROM read_parquet('{output}/small/*/*.parquet')")
+    conn.execute(f"CREATE OR REPLACE VIEW small AS SELECT * FROM read_parquet({small_parquets()})")
     # Deterministic coverage of each source, each task, and grounding format.
     sql="""SELECT * FROM small QUALIFY row_number() OVER (PARTITION BY source,task_type,input_modality ORDER BY selection_hash)=1"""
     rows=conn.execute(sql).fetch_arrow_table().to_pylist()

small/verification.json

@@ -8,14 +8,14 @@
   "missing_media": 0,
   "known_sha_split_overlap": 0,
   "rows": {
+    "train": 1000000,
     "validation": 50000,
-    "test": 50000,
-    "train": 1000000
+    "test": 50000
   },
   "groups": {
     "validation": 20709,
-    "train": 406008,
-    "test": 20733
+    "test": 20733,
+    "train": 406008
   },
   "task_coverage": {
     "validation": 20,