Dataset from Hugging Face · libredb
Database agent runs: 8,199 tool-calling runs across 40 language models
- License
- MITcommercial use allowedLicense text
- This license allows commercial use. Usually you must credit the author; read the license for the exact terms.
- Formats
- JSON, JSONL, PY, TXT
- Size
- 78 MB
- Updated
- Sep 21, 2026
Description
LibreDB Agent Benchmark 8,199 agent runs · 39 open-weight models served locally, plus one hosted model as a control · 6 task surfaces · 110,711 ledger events · 14,008 refused tool calls This is the complete measurement record behind the paper What Stops a Small Language Model From Driving a Database Agent. It is not a scored summary: it is every event the server wrote while the runs happened, released so that every number in the paper can be recomputed, and disagreed with…
Description from Hugging Face, shortened.
Files (8)
- JSONLevents.jsonl69 MB
- PYhf-export.py12 kB
- TXTleaderboard-pooled.txt2.5 kB
- TXTleaderboard-row.txt2.4 kB
- JSONLrefusals.jsonl4.4 MB
- JSONLruns.jsonl4.6 MB
- PYscore.py8.5 kB
- PYverify.py7.7 kB
DataHunter does not store or host these files: every link leads to the original source. Check the license there before using the data.