Spaces:
Configuration error
Configuration error
| name: fraudshield | |
| version: 0.2.0 | |
| description: | | |
| Production-grade OpenEnv environment for e-commerce fraud detection. | |
| Simulates marketplace fraud review with 3 difficulty levels (easy/medium/hard), | |
| deterministic graders, business-cost-sensitive rewards, and reproducible baselines. | |
| author: Devika J | |
| email: devikaj2005@gmail.com | |
| environment: | |
| class: FraudShieldEnvironment | |
| module: fraudshield_env | |
| action: | |
| class: FraudCheckAction | |
| module: models | |
| observation: | |
| class: FraudCheckObservation | |
| module: models | |
| reward: | |
| class: Reward | |
| module: models | |
| tasks: | |
| - name: easy | |
| description: | | |
| Clear-cut fraud detection with strong single-transaction indicators. | |
| Agents should achieve 95%+ accuracy by identifying obvious red flags. | |
| Difficulty: Foundation level - rewards strong pattern recognition. | |
| difficulty: easy | |
| num_transactions: 24 | |
| baseline_score: 1.0000 | |
| - name: medium | |
| description: | | |
| Mixed-signal reviews where confidence calibration and risk tradeoffs matter. | |
| No single feature is decisive; agents must weigh multiple indicators. | |
| Difficulty: Intermediate - rewards balanced decision-making. | |
| difficulty: medium | |
| num_transactions: 36 | |
| baseline_score: 0.8773 | |
| - name: hard | |
| description: | | |
| Coordinated fraud rings and legitimate flash-sale edge cases with overlapping signals. | |
| Complex patterns require network-level analysis and tolerance for ambiguity. | |
| Difficulty: Advanced - rewards sophisticated reasoning. | |
| difficulty: hard | |
| num_transactions: 48 | |
| baseline_score: 0.7206 | |
| metadata: | |
| created: "2026-03-30" | |
| framework: openenv | |
| license: MIT | |
| tags: | |
| - openenv | |
| - fraud-detection | |
| - e-commerce | |
| - agent-evaluation | |