Datasets:

samrat-kar
/

ap-sql-peft

license: apache-2.0
language:
  - en
task_categories:
  - text-generation
tags:
  - text-to-sql
  - oracle
  - accounts-payable
  - sql
  - chat
pretty_name: AP Text-to-SQL (Oracle) PEFT dataset
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: train
        path: train.jsonl
      - split: validation
        path: val.jsonl
      - split: test
        path: test.jsonl

ap-sql-peft

A chat-format text-to-SQL dataset for Accounts Payable analytics on Oracle. The schema is modelled on Oracle Fusion AP, Payments and Supplier tables, and the data behind it is synthetic. The dataset was used to train the LoRA adapter samrat-kar/ap-sql-v1.

Every gold SQL query was run against the demo database when the dataset was built; meta.result_rows records how many rows it returned.

Format

One JSON object per line:

{"messages": [{"role": "system", "content": "...rules + CREATE TABLE DDL + glossary..."},
              {"role": "user", "content": "question (or follow-up / repair request)"},
              {"role": "assistant", "content": "```sql\nSELECT ...\n```"}],
 "meta": {"id": "...", "tier": "T1-T5", "family": "...", "kind": "ask|followup|repair-exec|repair-parse",
          "source": "...", "group": "...", "tables": ["..."], "result_rows": 3, "approx_prompt_tokens": 3089, "split": "..."}}

The system prompt is the one the assistant builds at runtime: the rules, then the DDL of the retrieved tables (wide tables pruned to join keys plus the relevant columns), then the glossary entries. It is capped at about 3,584 tokens so the whole prompt fits a 4,096-token window.

Splits

tier	train	validation	test
T1 (single table, simple)	284	17	37
T2	463	28	56
T3	471	34	58
T4	143	10	19
T5 (hardest)	80	6	9
total	1,441	95	179

Split by meta.group. Paraphrases and variants of one question always land in the same split, so no test group appears in train.
Kinds. ask is a plain question. followup carries conversation history. repair-exec and repair-parse give the model a failed query and its Oracle or parse error, and the target is the corrected query.
Sources. S2-templated rows are generated from templates, S6-repair rows are repair cases, and S1-golden rows are hand-written (test split only).

Usage

from datasets import load_dataset
ds = load_dataset("samrat-kar/ap-sql-peft")

Limitations

Covers one synthetic schema.
Most rows are templated, so question phrasing is less varied than real user traffic.
Oracle SQL dialect only.