Agent training environment · 2026 · Top 10 · Meta x PyTorch
Agents learn to resolve invoice exceptions. Fine-tuning took the score from 0.14 to 0.73.
An OpenEnv environment where LLM agents work through invoices, purchase orders, goods receipts and policy rules to approve, hold, reject or escalate.
22 tasks with deterministic graders, so every run is reproducible. SFT and GRPO runs on Qwen with LoRA are scored against the same suite.