V4_bevel: bevel, the first modelling operation (README decision 46); mixes updated

.gitattributes

@@ -86,3 +86,4 @@ LT2_dialog/stress.jsonl filter=lfs diff=lfs merge=lfs -text
 LT2_dialog/train.jsonl filter=lfs diff=lfs merge=lfs -text
 V2_edits/train.jsonl filter=lfs diff=lfs merge=lfs -text
 V3_create/train.jsonl filter=lfs diff=lfs merge=lfs -text
+V4_bevel/train.jsonl filter=lfs diff=lfs merge=lfs -text

README.md

@@ -32,13 +32,14 @@ Training corpora for SpatialAi, a from-scratch model that understands and edits
 | `LT2_dialog/` | LT2 | 20,000 | 2,000 | 2,000 |
 | `V2_edits/` | V2 | 20,000 | 2,000 | 2,000 |
 | `V3_create/` | V3 | 20,000 | 2,000 | 2,000 |
+| `V4_bevel/` | V4 | 20,000 | 2,000 | 2,000 |
 
 Each folder holds `train.jsonl`, `val.jsonl` (held-out wordings and placements), `stress.jsonl` (turned, mirrored, restated and crowded scenes), `manifest.json` (seed, counts, settings) and `answers.json`. `mixes/` holds the weights that combine corpora into one run.
 
-Each build is tagged, so a run can pull exactly the data it trained on. Builds: `ladder-v1`, `ladder-v2`.
+Each build is tagged, so a run can pull exactly the data it trained on. Builds: `ladder-v1`, `ladder-v2`, `ladder-v3`.
 
 ```python
 from huggingface_hub import snapshot_download
-snapshot_download("khaines178/spatialai-corpora", repo_type="dataset", revision="ladder-v2",
+snapshot_download("khaines178/spatialai-corpora", repo_type="dataset", revision="ladder-v3",
                   allow_patterns=["S1_frames/*"], local_dir="data")
 ```

V4_bevel/answers.json

@@ -0,0 +1,12 @@
+{
+  "stage": "V4",
+  "case_sensitive": true,
+  "tasks": {
+    "ed.bevel": [],
+    "ed.bevel_width": [],
+    "ed.bevel_segments": [],
+    "ed.bevel_more": [],
+    "ed.bevel_set": [],
+    "ed.bevel_smooth": []
+  }
+}

V4_bevel/train.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c2822622d20443f96f390df47daf7e164c5579bb7ce4040baa941ff9668dc414
+size 76815854

mixes/mix_full_ladder3.json

@@ -86,15 +86,20 @@
       "weight": 2.0,
       "note": "V3: making and removing objects (README decision 45). Six tasks at the per-task share V2's edits have - unconfirmed (rule 4)"
     },
+    {
+      "path": "data/V4_bevel",
+      "weight": 2.0,
+      "note": "V4: bevel, the first modelling operation (README decision 46). Six tasks at the per-task share V2's and V3's have - unconfirmed (rule 4)"
+    },
     {
       "path": "data/R1_hssd",
-      "weight": 4.750071,
-      "note": "real rooms (HSSD, README decision 30) at 15% of the draw. The real share stays the user's 30% (2026-09-24), split evenly with recognition (the user's go, 2026-09-25). Re-weighted to 3.964286 on 2026-09-26 so real stays 30% with the whole ladder in the mix"
+      "weight": 5.178643,
+      "note": "real rooms (HSSD, README decision 30) at 15% of the draw. The real share stays the user's 30% (2026-09-24), split evenly with recognition (the user's go, 2026-09-25). Re-weighted to 3.964286 on 2026-09-26 so real stays 30% with the whole ladder in the mix. Re-weighted to 5.178643 on 2026-09-27 so real stays 30% with V4 in the mix"
     },
     {
       "path": "data/R2_objects",
-      "weight": 4.750071,
-      "note": "what a real object is, read through the geometry channel (README decision 32) - the other 15% of the real share. Needs model.geo_latents 4 and data.geo_library. Re-weighted to 3.964286 on 2026-09-26 so real stays 30% with the whole ladder in the mix"
+      "weight": 5.178643,
+      "note": "what a real object is, read through the geometry channel (README decision 32) - the other 15% of the real share. Needs model.geo_latents 4 and data.geo_library. Re-weighted to 3.964286 on 2026-09-26 so real stays 30% with the whole ladder in the mix. Re-weighted to 5.178643 on 2026-09-27 so real stays 30% with V4 in the mix"
     }
   ]
 }