---
language:
- en
license: apache-2.0
task_categories:
- text-generation
tags:
- sql
- text-to-sql
- enterprise
- unsloth
- chatml
- chain-of-thought
- postgresql
- snowflake
- clickhouse
- bigquery
- mysql
size_categories:
- n<1K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
  - split: regression
    path: data/regression-*
---

# 🚀 Enterprise Text-to-SQL & Analytical BI (Verbose CoT Reasoning)

This dataset contains **600 curated training records with in-depth, verbose 4-phase `<Thinking>` Chain-of-Thought reasoning**, **100 frozen evaluation benchmark samples**, and **50 frozen regression verification samples** formatted in standard **ChatML (`messages`)** and Prompt-Target pairs, strictly following the **Pioneer / Prometheus research paper** 3-slice curriculum design.

---

## 📊 Dataset Composition & 3-Slice Breakdown

| Split | Rows | Proportion | Paper Design & Purpose |
| :--- | :---: | :---: | :--- |
| **`train` (Slice 1: Gold CoT)** | **360** | **60.0%** | Full domain context + verbose 4-phase `<Thinking>` Chain-of-Thought reasoning (Constraint Analysis, Formal Derivation, Numerical Gating, Implementation). |
| **`train` (Slice 2: Hard Negatives)** | **180** | **30.0%** | Contrastive edge-case disambiguation & boundary refutations (e.g. Non-existent columns, TLE quadratic overhead, Regulation E APP fraud boundaries). |
| **`train` (Slice 3: Replay Buffer)** | **60** | **10.0%** | Domain primitives & optimization anchors to prevent catastrophic forgetting. |
| **Total `train` Split** | **600** | **100%** | Compact, high-signal, dense curated dataset for fine-tuning. |
| **`test` (Frozen Eval)** | **100** | — | Held-out benchmark with zero training contamination (`eval_leakage = 0`). |
| **`regression` (Baseline)** | **50** | — | Baseline regression validation set. |

---

## 🛡️ 5-Point Quality Audit (100% Passed)

* **Uniqueness**: `100.0%` unique prompts (600 / 600).
* **Zero Eval Leakage**: `0` prompt overlaps across train, test, and regression splits.
* **Replay Fraction**: `10.0%` dedicated to optimization and schema resilience.

---

## 🚀 Usage with Unsloth / Hugging Face

```python
from datasets import load_dataset

dataset = load_dataset("StarsMakeGalaxy/enterprise-text2sql-curated-600")
print(dataset)
```