Add real eval subset (300 real cable imgs) as 'real' config

README.md

@@ -19,6 +19,10 @@ configs:
   data_files:
   - split: train
     path: data/*.parquet
+- config_name: real
+  data_files:
+  - split: test
+    path: real/*.parquet
 ---
 
 # WireSeg-36K
@@ -45,6 +49,7 @@ masks and per-image metric depth. Single category: `cable`. Generated with **Def
 | path | what |
 |---|---|
 | `data/*.parquet` (config `default`) | RGB image + per-instance COCO-RLE masks + metadata, one row per image |
+| `real/*.parquet` (config `real`) | **real-image eval subset** (300 imgs) — RGB + COCO-RLE masks, one row per image |
 | `annotations.json` | full COCO ground truth (36,000 images / 176,387 annotations, single category `cable`) |
 | `depth/*.tar.gz` | per-image metric depth, float32 `[1024,1024]` `.npy`, packed in 100 gzipped tar shards |
 
@@ -80,10 +85,33 @@ d = np.load(f"depth/{ex['file_name'][:-4]}.npy")   # float32 [1024,1024], metric
 `jaccard_j`, `category`. Masks / bboxes / areas are JSON-encoded per-instance lists
 (RLE = COCO compressed).
 
+## Real eval subset (config `real`)
+
+A companion set of **300 real cable images** with hand-verified ground truth, for evaluating
+sim-to-real transfer of the synthetic `default` set. These were cherry-picked as the
+**disagreement slice** between base **SAM 3** (prompt `"cable"`) and a cable-tuned **LoRA** —
+i.e. the images the two models most disagreed on, so they concentrate the hard/informative
+cases. Masks are COCO-RLE, single category `cable`.
+
+- **Images:** 300 real (variable resolution) — `scene_family` = `real_easy` (224) / `real_hard` (76)
+- **Instances:** 2,119 cable masks (COCO RLE)
+
+```python
+from datasets import load_dataset
+real = load_dataset("DeformX/WireSeg-36K", "real", split="test")
+ex = real[0]                                  # ex["image"] -> PIL.Image
+```
+
+Fields: `image`, `file_name`, `scene_family`, `masks_rle`, `bboxes`, `areas`, `num_instances`,
+`category`, `width`, `height` — the mask/bbox/area columns match the `default` config (same
+JSON-encoded per-instance COCO-RLE), so decoding code is shared. The synthetic-only QA columns
+(`n_cable`, `wire_area_frac`, `det_*`, `jaccard_j`) are omitted; `width`/`height` are added
+because real images are not a fixed size.
+
 ## Splits
 
-A single `train` split is provided; users should split as needed (recommended: stratify by
-`scene_family` / `n_cable`).
+The `default` config provides a single `train` split; users should split as needed (recommended:
+stratify by `scene_family` / `n_cable`). The `real` config is a held-out `test` set.
 
 ## Generation
 

real/real-000.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d1fe5d890a441571bb1c921fcdcd3c0adc8da350173716265786b7ca930ce731
+size 474851243

real/real-001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f706479ed3f20c60bdf4569e427ca0452477e86c7e6973427d04a0bb998a14bb
+size 388957089

real/real-002.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b683e38328a70f40667dccb4cb4dd0d6c46f016792ba6b60e0a965672aa5bf52
+size 399646211