Remove fixed Matrix entity positions from training examples

Release v1.1.0: randomize Matrix entity row/column positions and HN exchange partners. Preserve the axis-breaking HN rule, all entity-pair answers and the default dataset schema. Synchronize title-based Matrix queries and all rendering metadata.

Constraint: Preserve all 1500 non-Matrix rows
Confidence: high
Scope-risk: moderate
Directive: Use a new version tag and append CHANGELOG.md for later releases
Tested: 2000 staged rows; 4000 PNGs; baseline non-Matrix equality; 500 Matrix answers and HN rules

CHANGELOG.md

@@ -0,0 +1,14 @@
+# Dataset version history
+
+## v1.1.0 — 20261109
+
+- Matrix table entity positions randomized independently across rows and columns.
+- HN swap rule preserved; valid exchange partner selected randomly.
+- Rerendered 500 Matrix gold/HN pairs (1,000 PNGs), seed 20261109.
+- Synced title-based Matrix queries, HN records, rendered specs and provenance.
+- Preserved all 1,500 non-Matrix rows and the default 2,000-row train split.
+- Parent revision: `2d4602a15971e1d9a032f730c53dbaef33c0a267`.
+
+## v1.0.0 — baseline
+
+Original dataset before this release: `2d4602a15971e1d9a032f730c53dbaef33c0a267` (2,000 pairs).

README.md

@@ -25,17 +25,16 @@ dataset_info:
     dtype: string
   splits:
   - name: train
-    num_bytes: 430733374.0
+    num_bytes: 430784168.0
     num_examples: 2000
-  download_size: 427805046
-  dataset_size: 430733374.0
+  download_size: 427785946
+  dataset_size: 430784168.0
 configs:
 - config_name: default
   data_files:
   - split: train
     path: data/train-*
 ---
-
 # Table retrieval training data
 
 2,000 synthetic hard table retrieval pairs generated with GPT-6 Luna: 500 each of Vertical, Matrix, Hierarchy, and Nested.
@@ -50,3 +49,16 @@ Negative rules: Vertical swaps distinct values in one column; Matrix moves the q
 from datasets import load_dataset
 ds = load_dataset("NLPRETGOOD/table_train", split="train")
 ```
+
+## Version v1.1.0: randomized Matrix entity positions
+
+Matrix 엔티티의 행·열 위치를 무작위화했습니다. Gold tables independently shuffle rows and columns while preserving every entity-pair value. Hard negatives keep the same axis-swap rule and select their valid exchange partner randomly. All 500 Matrix pairs (1,000 PNGs) were rerendered with seed `20261109`; Matrix queries now use the visible table title. The remaining 1,500 training rows are unchanged.
+
+See [CHANGELOG.md](CHANGELOG.md) and [versions/v1.1.0.json](versions/v1.1.0.json) for source image hashes, position counts, provenance and the baseline revision. Future releases use a new immutable `vX.Y.Z` tag and append their changes to the changelog.
+
+Pin this version for reproducible training:
+
+```python
+from datasets import load_dataset
+ds = load_dataset("NLPRETGOOD/table_train", revision="v1.1.0", split="train")
+```

VERSION

@@ -0,0 +1 @@
+v1.1.0

data/train-00000-of-00003.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7183dfd43df3c19ca936c9c674ac907b510e2f2efb1d4b0567cb0cec20c8cbd2
-size 95718408
+oid sha256:c3162a5b6633a2ddad7dc35cdcabe698fe0f2c3b64a597fd7604b5f0c01f7046
+size 95721373

data/train-00001-of-00003.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6f47dfa175405f3dad39c157977b0f13b99c192cb8af55c93d2c97cfcd1d5570
-size 108138498
+oid sha256:78299be99f7731540c37c1f81ed8a561486d23fd455c2cebd8b884ac11158932
+size 108116433

metadata/rendered_assets/hard/gpt-6-luna-hard-train-500/generation_manifest.json

@@ -0,0 +1,18 @@
+{
+  "source": "/data/hoseob/llm/Project/vc_final/generated_content_train/hard/gpt-6-luna-hard-train-500",
+  "total_count": 2000,
+  "table_types": {
+    "vertical": 500,
+    "matrix": 500,
+    "hierarchy": 500,
+    "nested": 500
+  },
+  "format": "png",
+  "style": "random",
+  "seed_by_type": {
+    "vertical": 20261108,
+    "matrix": 20261109,
+    "hierarchy": 20261110,
+    "nested": 20261111
+  }
+}