Team: Spektr-SQL
Affiliation: Amazon Ads - SpektrBot
Authors: Kishore Kumar, Scott Roberts, Desik Rengarajan, Ravi Shankar (Amazon Ads)
Submitted: June 15, 2026
Category: 4 (Combined Models — GPU + API)
| Dev (1531 questions) | Test | |
|---|---|---|
| Execution Accuracy | 73.09% (1119/1531) | 78.31% (1401/1789) |
| Database | Accuracy | Questions |
|---|---|---|
| superhero | 90.7% | 117/129 |
| student_club | 86.7% | 137/158 |
| codebase_community | 76.1% | 140/184 |
| european_football_2 | 76.0% | 98/129 |
| debit_card_specializing | 75.0% | 48/64 |
| california_schools | 73.0% | 65/89 |
| financial | 72.6% | 77/106 |
| toxicology | 70.3% | 102/145 |
| card_games | 64.7% | 123/190 |
| formula_1 | 64.9% | 113/174 |
| thrombosis_prediction | 60.7% | 99/163 |
Two-Phase Agentic Ensemble Pipeline:
Phase 1 (GPU): Fine-tuned Qwen3-30B-A3B MoE model generates 36 SQL candidates per question.
Phase 2 (API): Claude Sonnet 4.6 runs a multi-turn agentic verification loop with database exploration tools. If it disagrees with the FT, Claude Opus 4.6 acts as tiebreaker.
| Component | Specification |
|---|---|
| GPU | 8× H100 80GB |
| GPU Model | Qwen3-30B-A3B-Instruct-2507 (MoE, ~3B active) |
| API Models | Claude Sonnet 4.6 + Claude Opus 4.6 |
| Total Runtime | ~8-10 hours |
| API Cost | ~$200-250 (with prompt caching) |
| Column Meaning | YES |
| External Knowledge | YES |
| Self-Consistency | Many (n=36) |