Use cleaned TopBench data on main

Make main the default dataset download and remove obsolete release links. All task data and oracles are unchanged.

CHANGELOG.md

@@ -2,7 +2,7 @@
 
 ## cleaned-20261107
 
-All 779 questions, 292 dataset folders, and 214 `test_current.csv` oracles are retained. Question IDs and the inference/evaluation JSON structure are unchanged.
+The `main` branch provides the current cleaned data. All 779 questions, 292 dataset folders, and 214 `test_current.csv` oracles are retained. Question IDs and the inference/evaluation JSON structure are unchanged.
 
 - Corrected reviewed qualitative/categorical descriptions, codebook interpretations, feature aliases, outcome proxies, and ranking direction metadata.
 - Applied strict visible-input group holdout to seven hotel batches: removed 387 overlapping history rows, without changing current candidates or outcome labels.
@@ -22,8 +22,6 @@ Historical model outputs and summary scores are not results for this cleaned ver
 
 Deterministic validation and full-dataset automatic semantic screening, followed by a second pass over flagged cases, were supplemented by source-based adjudication. This is not exhaustive human annotation or a guarantee that every qualitative phrase is error-free. Qualitative language has no published one-to-one quantile mapping. Treatment Effect Analysis is predictive scenario comparison rather than causal identification.
 
-The previous public dataset remains available at revision [`643090b74d9beb0da9f2da3ff650af70c12511f1`](https://huggingface.co/datasets/LAMDA-Tabular/TopBench/tree/643090b74d9beb0da9f2da3ff650af70c12511f1).
-
 ### Character-Restoration Sources
 
 - [Forest Fires in Brazil: original Kaggle data](https://www.kaggle.com/datasets/gustavomodelli/forest-fires-in-brazil) (Latin-1 archive; not the already-lossy local UTF-8 copy).

DATA_MANIFEST.json

@@ -8,18 +8,18 @@
   "files": [
     {
       "path": "CHANGELOG.md",
-      "bytes": 3890,
-      "sha256": "9fab577c44e522ce6edf82d05813c8c911cb23d04b2df347454a0cfb50b49d60"
+      "bytes": 3736,
+      "sha256": "c8b298f19c5f0292528102079e7052c434daa4645ab5a04390d18b1cbc40e795"
     },
     {
       "path": "README.md",
-      "bytes": 3989,
-      "sha256": "45290a441d08308e0c0dd7426ca2f42dc3b50d2281de26607b0e9c20d5e28aea"
+      "bytes": 3865,
+      "sha256": "f3284b9694ec5e9d0fb63741c3523899521227f6c481eb4c110959f7ddfdd131"
     },
     {
       "path": "data_version.json",
-      "bytes": 769,
-      "sha256": "502e2ce36f4c41d0d27186e3c8844fc8f7c1db407d32254ad2afbcaa163452d6"
+      "bytes": 725,
+      "sha256": "44fd52e5b5ca3dec7dd81cc8c43e17206aa6016cb7a14bd931b6205aa370aa77"
     },
     {
       "path": "decision_making/daily/classification/Apple_Quality_B2/Apple_Quality_B2_001.json",

README.md

@@ -32,10 +32,11 @@ TopBench evaluates implicit predictive reasoning from natural-language questions
 
 ## Download
 
+The `main` branch contains the current cleaned dataset.
+
 ```bash
 pip install -U huggingface_hub
-hf download LAMDA-Tabular/TopBench --repo-type dataset \
-  --revision cleaned-20261107 --local-dir data
+hf download LAMDA-Tabular/TopBench --repo-type dataset --local-dir data
 ```
 
 See the [code README](https://github.com/LAMDA-Tabular/TopBench#readme) for environment setup, sandbox configuration, inference, and evaluation. Use the downloaded `data` directory as `--data-root`.
@@ -50,9 +51,9 @@ Task folders contain `history.csv`, question JSON files, and `info.json` / `info
 
 ## Data Version
 
-`cleaned-20261107` is an audited data update, **not the original paper-score dataset**. It corrects documented query mappings, target aliases, ranking scope and completeness, corrupted category text, and hotel visible-input group overlap. All 779 question slots and 214 oracle files are retained.
+`main` currently provides data version `cleaned-20261107`, an audited data update, **not the original paper-score dataset**. It corrects documented query mappings, target aliases, ranking scope and completeness, corrupted category text, and hotel visible-input group overlap. All 779 question slots and 214 oracle files are retained.
 
-For the previous public data, pin revision [`643090b74d9beb0da9f2da3ff650af70c12511f1`](https://huggingface.co/datasets/LAMDA-Tabular/TopBench/tree/643090b74d9beb0da9f2da3ff650af70c12511f1). Re-run inference and evaluation on this version in a fresh output directory; previous model outputs and scores are not interchangeable with the cleaned data. See [CHANGELOG.md](CHANGELOG.md) and [data_version.json](data_version.json).
+Use the current dataset for inference and evaluation in a fresh output directory; previous model outputs and scores are not interchangeable with the cleaned data. See [CHANGELOG.md](CHANGELOG.md) and [data_version.json](data_version.json).
 
 Checks cover file integrity, loaders, ranking oracles, observed targets, source-supported repairs, and full-dataset automatic semantic screening with a second pass over flagged cases and evidence-based adjudication. They are not exhaustive independent human annotation. Qualitative descriptions remain approximate rather than invertible numeric encodings. Treatment Effect Analysis uses predictive comparisons of observed scenarios, not identified causal effects; auxiliary fields in source records may differ beyond the attributes stated in a question.
 

data_version.json

@@ -2,7 +2,6 @@
   "version": "cleaned-20261107",
   "release_date": "2026-11-07",
   "repo": "LAMDA-Tabular/TopBench",
-  "previous_public_revision": "643090b74d9beb0da9f2da3ff650af70c12511f1",
   "total_questions": 779,
   "question_counts": {
     "decision_making": 186,
@@ -15,5 +14,6 @@
   "score_compatibility": "Not interchangeable with historical model outputs or paper scores; re-evaluate separately",
   "changelog": "CHANGELOG.md",
   "manifest": "DATA_MANIFEST.json",
-  "validation_scope": "Deterministic integrity and source checks; automatic semantic review plus evidence-based adjudication, not exhaustive human annotation"
+  "validation_scope": "Deterministic integrity and source checks; automatic semantic review plus evidence-based adjudication, not exhaustive human annotation",
+  "default_revision": "main"
 }