Datasets:
File size: 3,617 Bytes
0be14c9 702474b 0be14c9 702474b 0be14c9 d56e423 0be14c9 702474b 0be14c9 702474b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 | ---
license: mit
pretty_name: "EdTech SQL Evaluation (declared learning curve, verified answer key)"
tags:
- synthetic
- text-to-sql
- evaluation
- data-agents
- tabular
- sql
size_categories:
- 10K<n<100K
task_categories:
- table-question-answering
configs:
- config_name: students
data_files: "tables/students.csv"
- config_name: practice_attempts
data_files: "tables/practice_attempts.csv"
---
# EdTech SQL Evaluation (declared learning curve, verified answer key)
An **evalpack**: an evaluation database generated from the answer key, not
annotated after the fact. A VLDB 2026 audit found 52.8% of BIRD Mini-Dev
answer keys wrong because benchmarks annotate answers onto existing
databases; this dataset inverts the order. The declared properties (curves,
shares, identities) are the specification, the database is generated to
satisfy them exactly, and every shipped question was re-verified by
executing its gold SQL against these exact files with DuckDB, an engine
that shares no code with the generator.
**15,600 rows** across the tables below.
## What is in it
Tables: students (600 rows), practice_attempts (15,000 rows).
A practice log where the monthly correct rate climbs 55% to 68% by
declaration and total practice time per month is declared and exact. Every
attempt belongs to a student who exists and happens after that student
signed up. 17 questions ship; 4 candidates whose rates were unachievable at
the generated row counts were dropped by the verification gate rather than
shipped wrong (that refusal is the point).
## Files
- `tables/*.csv`: the database
- `questions.jsonl`: one verified question per line (natural language,
gold SQL, expected answer, tags)
- `certificate.json`: per-question DuckDB verification plus FK proof
- `manifest.json`: spec hash, seed, library versions, dropped candidates
- `verify.py`: standalone re-verification (needs only `pip install duckdb`)
- `schema.misata.yaml`: the full declaration; regenerate or rotate the pack
from it
## Re-verify in thirty seconds
```bash
pip install duckdb
python verify.py
```
## Rotate the environment without touching the answer key
```bash
pip install misata
misata evalpack --config schema.misata.yaml -o rotated_pack --seed 7
```
A new seed replaces the rows; the declared answers stay the answers, so
agents cannot memorize the environment. Same version + same schema + same
seed reproduces these bytes exactly.
## Provenance
No real data: fully declaration-generated, offline, deterministic
([provenance statement](https://github.com/rasinmuhammed/misata/blob/main/DATA-PROVENANCE.md)).
Generated with [misata](https://github.com/rasinmuhammed/misata) (MIT).
Where the approach fails is documented in
[LIMITATIONS.md](https://github.com/rasinmuhammed/misata/blob/main/LIMITATIONS.md).
## The rest of the shelf
Same idea, other domains, every answer re-verified with DuckDB against the
shipped files:
- [Ecommerce analytics](https://huggingface.co/datasets/rasinmuhammed/ecommerce-analytics-sql-evaluation)
- [SaaS finance](https://huggingface.co/datasets/rasinmuhammed/saas-finance-sql-evaluation)
- [EdTech practice](https://huggingface.co/datasets/rasinmuhammed/edtech-sql-evaluation)
If you want ground truth for a regression rather than for SQL, the
[machine degradation set](https://huggingface.co/datasets/rasinmuhammed/predictive-maintenance-remaining-useful-life)
carries exact remaining useful life on 100 run-to-failure machines.
Generated by [Misata](https://github.com/rasinmuhammed/misata) (MIT).
Build your own: [misata.studio](https://www.misata.studio).
|