Reward Atlas
All field guidesEsc to close
Sign in
LLM / AgenticPublicPublishedv1.0Advanced

SQL-Detective-v1

A multi-turn agent answers analytics questions by querying a read-only SQLite sandbox. Tool use, verifiable rewards, 12-turn budget. Maintained by Bhaskar & Jaswanth · updated 28 Aug 2026.

Dataset
2,400 q · 12 DBs
Action
tool call | final answer
Turn cap
12
Best baseline
GRPO-7B 61.4%
05

Reward function

Rubric

Reward is given once, when submit_answer is called. There is no per-turn shaping.

scorecriterionnote
1.0Exact result-set matchOrder-insensitive multiset comparison
0.2Partial column match≥ 50% of expected columns present with correct values
−0.1Per invalid query after the 3rdFloors at 0
0Format canary triggeredAnswer contains the hidden canary tokens
Terminated

submit_answer is called. The rubric is scored and the episode ends.

Truncated

12 turns elapse without an answer. Reward 0, with no terminal bonus.

Reward-hacking checks
  • format canaryPASS
  • answer without queryingPASS
  • SELECT * dumpFAIL

    Submitting a full table can reach 0.2 partial credit.

Guide details
Version
Type
LLM / Agentic
API
OpenEnv
License
Apache-2.0
Seeds
3
Domains
tool-usedatacoding
Install
pip install reward-atlas-sql-detective

Issue with this step? Suggest an edit