thesis-split: 784/196 partition stratified by tier, spatial function and territorial division

README.md

@@ -46,10 +46,28 @@ The dataset comprises manually annotated question-SQL pairs classified into four
 
 ## 📁 Files and Reproducibility
 
-The dataset is divided into standard `train.parquet` and `validation.parquet` splits. 
+The dataset is divided into standard `train.parquet` and `validation.parquet` splits. Two tagged snapshots are preserved so that published results remain reproducible:
+
+| Tag | Content | Used in |
+| :--- | :--- | :--- |
+| `paper-released` | 980 curated pairs (`source = base_dataset`) plus five augmented partitions in `train.parquet`; a separate 196-pair `validation.parquet` | SBBD 2026 paper (first execution of the experiments) |
+| `thesis-split` (current `main`) | The 980 curated pairs partitioned into 784 training and 196 validation pairs, stratified by complexity tier, spatial function and territorial division (one validation pair per tier × function × division cell, seed 42); `split_manifest.json` records the counts and checksums | Master's thesis (second execution of the experiments) |
+
+### Loading the thesis split (Python)
+
+```python
+from datasets import load_dataset
+
+dataset = load_dataset("datafromlopes/atlas-sql-br", revision="thesis-split")
+print(len(dataset["train"]), len(dataset["validation"]))   # 784 196
+```
+
+The thesis split is generated by `core/make_paper_split.py` in the [code repository](https://github.com/datafromlopes/atlas-sql-br), which also holds the training, generation and evaluation pipeline and the results of the thesis run. The fine-tuned adapter is published as [`datafromlopes/llama-3.1-8b-atlassql-br`](https://huggingface.co/datafromlopes/llama-3.1-8b-atlassql-br).
+
+### Reproducing the paper
 
 **⚠️ Important Note on Reproducibility:**
-The baseline experiments documented in the original paper were conducted **exclusively** on the original 980 question-SQL pairs. The dataset files available here also include additional data generated through data augmentation techniques. 
+The baseline experiments documented in the original paper were conducted **exclusively** on the original 980 question-SQL pairs. The `paper-released` snapshot also includes additional data generated through data augmentation techniques. 
 
 To exactly replicate the experiments and results from the original paper, you **must**:
 1. Use the dataset version associated with the **`paper-released`** tag.

split_manifest.json

@@ -0,0 +1,68 @@
+{
+  "source_file": "/Users/diegolopes/repositories/atlas_sql_br/data/paper_released/train.parquet",
+  "source_md5": "bd62f326098f882009ca72b7f26c5c97",
+  "source_partition": "base_dataset",
+  "seed": 42,
+  "val_fraction": 0.2,
+  "stratified_by": [
+    "level",
+    "geospatial_functions",
+    "territorial_division"
+  ],
+  "train": {
+    "rows": 784,
+    "by_tier": {
+      "Difícil": 196,
+      "Fácil": 196,
+      "Muito Difícil": 196,
+      "Médio": 196
+    },
+    "by_function": {
+      "ST_Buffer": 112,
+      "ST_Contains": 112,
+      "ST_DWithin": 112,
+      "ST_Distance": 112,
+      "ST_Intersection": 112,
+      "ST_Intersects": 112,
+      "ST_Within": 112
+    },
+    "by_division": {
+      "Bairro": 112,
+      "Distrito": 112,
+      "Estado": 112,
+      "Município": 112,
+      "País": 112,
+      "Região": 112,
+      "Rua": 112
+    }
+  },
+  "validation": {
+    "rows": 196,
+    "by_tier": {
+      "Difícil": 49,
+      "Fácil": 49,
+      "Muito Difícil": 49,
+      "Médio": 49
+    },
+    "by_function": {
+      "ST_Buffer": 28,
+      "ST_Contains": 28,
+      "ST_DWithin": 28,
+      "ST_Distance": 28,
+      "ST_Intersection": 28,
+      "ST_Intersects": 28,
+      "ST_Within": 28
+    },
+    "by_division": {
+      "Bairro": 28,
+      "Distrito": 28,
+      "Estado": 28,
+      "Município": 28,
+      "País": 28,
+      "Região": 28,
+      "Rua": 28
+    }
+  },
+  "train_md5": "c759e6c383ed39459ea9622dc1160911",
+  "validation_md5": "645ff583d3deb55dbbe02807d6e0ecaf"
+}

train.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a1e32457ee55f7577842fa55a0213e73bb5e4af9534430b683012f28ce5e6ffe
-size 228093
+oid sha256:e20374824c82fa5fb66e4a2c6bcf2a807c61c098c6c36aea569638317e1d7c1c
+size 92006

validation.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:98fb2bb335aa4277894644b47f0beca9554bc55d4dc0eee9550c414ffc919271
-size 18772
+oid sha256:2e4c9af1dd118b7b7409200002f93597dd3e83f318580de1561d6abb7884fb38
+size 32722