Release TableSuite-1K v2.1.0

README.md

@@ -32,14 +32,6 @@ configs:
         path: prediction_episodes/validation/*.parquet
       - split: test
         path: prediction_episodes/test/*.parquet
-  - config_name: grounding_tasks
-    data_files:
-      - split: train
-        path: grounding_tasks/train/*.parquet
-      - split: validation
-        path: grounding_tasks/validation/*.parquet
-      - split: test
-        path: grounding_tasks/test/*.parquet
   - config_name: table_grounding
     data_files:
       - split: train
@@ -70,34 +62,40 @@ configs:
 
 # TableSuite-1K
 
-TableSuite-1K is a source-grounded benchmark for predictive and
-language-grounded tabular intelligence over 1,000 OpenML-referenced datasets.
-It provides source metadata and value-free task plans, not copied OpenML
-tables, model outputs, or checkpoints.
+TableSuite-1K benchmarks predictive and language-grounded tabular intelligence
+over 1,000 OpenML-referenced datasets.
 
-## Capabilities
+| Task | Input | Evaluation |
+| --- | --- | --- |
+| Prediction | ICL rows or a partially labelled serialized table | classification and regression |
+| Table grounding | a provided table plus a lookup/comprehension question | exact displayed-table facts |
+| Table QA | a provided subtable plus a typed question | programmatic operations |
+
+This repository contains metadata and value-free task plans. It contains no
+OpenML source values, labels, rendered questions, gold answers, model outputs,
+embeddings, or checkpoints. OpenML remains the source-table distributor.
+
+## Configurations
 
 | Configuration | Purpose |
 | --- | --- |
-| `datasets` | source, schema, target, split, and license metadata |
-| `table_prediction_tasks` | prediction eligibility and metrics |
-| `prediction_episodes` | frozen 4/16/32-shot support/query references |
-| `grounding_tasks` | eligible non-target columns |
-| `table_grounding` | provided-table cell/row/column operations |
-| `table_question_answering` | programmatic aggregate and lookup operations |
-
-Questions and gold answers are generated deterministically from the displayed
-source slice at access time. No LLM authors benchmark gold.
+| `datasets` | source identity, schema, target, split, and provenance |
+| `table_prediction_tasks` | prediction eligibility and primary metrics |
+| `prediction_episodes` | frozen prediction query anchors |
+| `table_grounding` | official provided-table grounding plans |
+| `table_question_answering` | official programmatic QA plans |
 
 ## Quickstart
 
+Install the matching package and materialize only the OpenML sources you need:
+
 ```bash
 pip install \
-  'tablesuite[local,hf,openml] @ git+https://github.com/Sichao-Li/TableSuite-1K.git@v2.0.1'
+  'tablesuite[local,hf,openml] @ git+https://github.com/Sichao-Li/TableSuite-1K.git@v2.1.0'
 
 tablesuite fetch-openml \
   --reference Lester1996/TableSuite-1K \
-  --revision v2.0.0 \
+  --revision v2.1.0 \
   --output openml-parquet \
   --dataset-id openml_45069 \
   --accept-source-terms
@@ -109,42 +107,52 @@ from tablesuite import TableSuite
 suite = TableSuite.open(
     "Lester1996/TableSuite-1K",
     source="openml-parquet",
-    revision="v2.0.0",
+    revision="v2.1.0",
 )
+
 task = suite.official(
     "table_grounding",
     split="dataset_test",
     dataset_ids=("openml_45069",),
 )
-
 example = task[0]
-result = task.score(example.id, model(example.prompt))
+score = task.score(example.id, model(example.prompt))
 ```
 
-## Protocol Summary
+Prediction uses the same frozen queries across interfaces and a deterministic
+nested support schedule:
+
+```python
+prediction = suite.prediction(
+    "icl",
+    support=(0.0, 0.1, 0.3, 0.5, 0.7, 0.9, 1.0),
+    dataset_ids=("openml_45069",),
+)
+```
 
-Prediction supports zero-shot and 4/16/32-shot row ICL, plus zero-label and
-partially labelled serialized-table inference. Serialized tables can cover the
-full source table or the same frozen query rows as the matched ICL protocol.
-It is inference-only. Every request states its task family, and classification
-requests list the allowed target labels without revealing row-level targets.
+For model context limits, `PredictionDataset.fit_context` selects the largest
+support prefix that fits the actual tokenized prompt and emits an auditable
+coverage report. It never silently truncates.
 
-Table grounding operates only on the table shown in the request. Official v2
-operations are `cell_lookup`, `row_lookup`, `column_values`,
-`distinct_values`, and `value_counts`.
+## Evaluation Contracts
 
-Table QA uses independently balanced 4/8/16-row subtables and typed count,
-sum, mean, min, max, argmax, and filter-then-argmax operations.
+- Prediction is inference-only; no per-dataset parameter updates are allowed.
+- Query targets are always private; only selected support labels are visible.
+- Grounding and QA operate only on the displayed table slice.
+- Wording and gold are generated deterministically from local source data.
+- Dataset transfer uses duplicate-aware `dedup_cluster_id` partitions.
+- Results must report response coverage and, for prediction, requested/realized
+  support plus context coverage.
 
-The v2 semantic tasks use literal source-schema wording. This release does not
-claim curated cross-dataset ontology equivalence.
+The v2.1 semantic tasks use literal source headers. This release does not claim
+curated cross-dataset ontology equivalence.
 
 ## Source Terms
 
-OpenML remains the source-table distributor. Each dataset retains its upstream
-terms. `openml_license_claim` is provenance metadata, not a license granted by
+Each referenced OpenML dataset retains its upstream terms.
+`openml_license_claim` is provenance metadata, not a license granted by
 TableSuite-1K. The repository-level `other` designation reflects heterogeneous
 source terms.
 
-Code and complete protocol documentation:
+Code and full protocol documentation:
 <https://github.com/Sichao-Li/TableSuite-1K>

datasets/test/part-00000.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:784cb040b679b6139ac5d151adb2f625ae4d99e3888680361ad8ded948cc3796
-size 22941
+oid sha256:f40dabc581c8ebf81ff8e564bd990e9438b878c6d53aa9cba36060fae0c1aac8
+size 34731

datasets/train/part-00000.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2c354652d55b98ff895fd893fdb281c9dcd9e73c4e1b1bed983054e7978f84da
-size 85958
+oid sha256:69a2f00be0c8f906ec4b36e335f571da9717fb7b8e280ebadc384bebc877a98a
+size 137255

datasets/validation/part-00000.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:accc08b39372954e0b30504cb7f4e71d245784218032c3af2558db5715f3fcd6
-size 20158
+oid sha256:b66236aaa8bef7b2ad75e9061523d02135f2d05b0e2d98d0b8ac6e52782b1aa8
+size 30029

grounding_tasks/test/part-00000.parquet

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:9f708de3fd82a799454c1aeabd37a851b094f30ffb0b1bea029131611e1049f6
-size 14165

grounding_tasks/train/part-00000.parquet

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:fb38662c641cb33a4fda0a09f6ede3b7d36e7d482520e7154c83976296b05771
-size 57128

grounding_tasks/validation/part-00000.parquet

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:b7da06ed78462bbd3b1f639cb0898ef2afe98c114271e4629791172cace550ce
-size 11891

reference_summary.json

@@ -23,29 +23,6 @@
         }
       }
     },
-    "grounding_tasks": {
-      "records": 989,
-      "splits": {
-        "test": {
-          "records": 96,
-          "shards": [
-            "test/part-00000.parquet"
-          ]
-        },
-        "train": {
-          "records": 805,
-          "shards": [
-            "train/part-00000.parquet"
-          ]
-        },
-        "validation": {
-          "records": 88,
-          "shards": [
-            "validation/part-00000.parquet"
-          ]
-        }
-      }
-    },
     "prediction_episodes": {
       "records": 35472,
       "splits": {
@@ -95,11 +72,11 @@
   },
   "contains_source_values": false,
   "record_schemas": {
-    "catalog": "1.0",
+    "catalog": "1.1",
     "official_tasks": "2.0"
   },
   "reference_id": "tablesuite-1k:2.0",
-  "release_version": "2.0.0",
+  "release_version": "2.1.0",
   "schema_version": "2.0",
   "source_provider": "openml"
 }