Request to add sriram1983007/sra-stablecoin-risk-bench to the benchmark allow-list

A Hugging Face user submitted the sriram1983007/sra-stablecoin-risk-bench dataset for inclusion in the platform's benchmark allow-list, asking the Hugging Face team to add it per the Evaluation Results documentation. The benchmark measures how models handle stablecoin (USDC) payments made by AI agents across two held-out test tasks: risk_gate (1,175 cases, including 69 prompt injections, where a model approves, holds or rejects a payment) and dispute (825 cases, where a model picks the remedy possible given where a payment sits relative to settlement finality). The dataset uses the inspect-ai evaluation framework with eval_risk_gate and eval_dispute configs, temperature 0 generation, a pattern scorer on the decision or outcome field, a full scorer in score.py covering macro-F1, unsafe approvals, wrongful refunds and denials, and is licensed CC-BY-4.0.

Hi Hugging Face team, The Evaluation Results docs say to get in touch to add a dataset to the benchmark allow-list, so I’m asking here. Dataset eval.yaml is in the repo root : sriram1983007/sra-stablecoin-risk-bench · Datasets at Hugging Face https://huggingface.co/datasets/sriram1983007/sra-stablecoin-risk-bench What it measures: how a model handles stablecoin USDC payments made by AI agents. Two tasks, both with held-out test splits: - risk gate 1,175 cases : approve, hold or reject a payment, given the policy, the payment request treated as untrusted, including 69 prompt injections and the results of verification tools. - dispute 825 cases : choose the remedy that is actually possible given where the payment sits relative to settlement finality. Setup: - eval.yaml uses evaluation framework: inspect-ai, with configs eval risk gate and eval dispute, split test, a system-message solver with the task prompt, generate at temperature 0, and a pattern scorer on the decision or outcome field. - The full scorer macro-F1, unsafe approvals, wrongful refunds and denials is score.py in the same repo. - License CC-BY-4.0.