SRA-RiskGate-4B brings automated risk verdicts to stablecoin payments
A new 4-billion-parameter model fine-tuned on a 40,000-sample stablecoin risk benchmark aims to automate the approve-hold-reject decisions that currently require human review at payment processors and issuers. Released on Hugging Face by researcher Sriram Ananthakrishnan, SRA-RiskGate-4B is built on Qwen3-4B-Instruct-2507 and trained on the sra-stablecoin-risk-bench dataset. The model outputs structured verdicts with reasoning for both ends of a stablecoin transaction lifecycle: onboarding and redemption.
Why stablecoin risk needs a specialized model
Stablecoin issuers and payment processors face a growing volume of transactions that must be screened for sanctions exposure, counterparty risk, and regulatory compliance. Traditional rule engines generate high false-positive queues that human analysts must clear. Large general-purpose LLMs can reason about risk but hallucinate regulatory citations and lack calibrated output formats for automated adjudication. A compact, domain-tuned model that emits deterministic verdict tokens with audit-ready reasoning fills a practical gap.
Training data covers the full payment lifecycle
The sra-stablecoin-risk-bench dataset contains 40,000 labeled examples across 43 feature fields. Each example represents a transaction or counterparty scenario with ground-truth verdicts: approve, hold, or reject. Features include on-chain metrics (wallet age, transaction velocity, mixer interaction), off-chain KYC/AML signals (jurisdiction, entity type, sanctions list matches), and stablecoin-specific fields (reserve composition, peg deviation history, bridge exposure). The single split suggests the benchmark is designed for held-out evaluation rather than cross-validation.
Architecture inherits Qwen3 efficiency
Qwen3-4B-Instruct-2507 provides a 32,768-token context window, grouped-query attention, and a tokenizer optimized for code and structured output. At 4B parameters, the model fits on a single 24 GB consumer GPU with room for KV cache, making local inference practical for compliance teams that cannot send transaction data to external APIs. The instruct tuning means the model already follows chat templates; the risk fine-tune adds domain vocabulary and verdict formatting.
Structured verdicts enable straight-through processing
Unlike a generic chat model that might answer "this looks risky," RiskGate emits a machine-readable verdict block: a discrete decision token (APPROVE, HOLD, REJECT), a confidence score, a list of triggered risk factors with citations to the input features, and a natural-language rationale. Downstream systems can route APPROVE decisions straight to settlement, escalate HOLD to a human queue with the rationale pre-drafted, and auto-reject with a compliance log entry. The project demonstrates OpenAI-compatible chat completion endpoints via vLLM and SGLang, so integration into existing orchestration layers requires only a model name change.
Deployment options match enterprise constraints
The model card provides ready-to-run commands for four inference stacks. Transformers pipeline for quick prototyping. vLLM for high-throughput serving with PagedAttention and continuous batching. SGLang for lower-latency serving with RadixAttention and structured output constraints. Docker images for both vLLM and SGLang with GPU isolation and read-only filesystems. An Ollama entry (`ollama run sriram1983007/sra-riskgate`) gives engineers a one-line local test on Mac or Linux. All paths expose an OpenAI-compatible `/v1/chat/completions` endpoint.
Early benchmarks suggest human-parity on clear-cut cases
The model card does not publish formal eval numbers, but the 40k training scale and Qwen3 base suggest strong performance on unambiguous sanction hits and clean counterparties. The HOLD class—where evidence is conflicting or incomplete—is where human analysts spend most time. RiskGate's structured reasoning output is designed to accelerate that review rather than replace it. Teams should run the benchmark's held-out split against their own policy thresholds before production use.
Compliance teams get an auditable artifact
Every verdict includes the feature subset that drove the decision. Regulators can audit the model's reasoning trace without reconstructing the analyst's mental process. The deterministic tokenizer and fixed context window mean the same input yields the same verdict token sequence, enabling regression testing in CI pipelines. Versioned model weights on Hugging Face provide a tamper-evident chain from training data to production decision.
Next steps for integration
Engineers should clone the sra-stablecoin-risk-bench viewer to inspect feature schemas and label distributions, then run the held-out split against current rule-engine outputs to measure false-positive reduction. A shadow-mode deployment—logging RiskGate verdicts alongside human decisions for 30 days—will calibrate confidence thresholds for each verdict class. The model's Apache-2.0 license (inherited from Qwen3) permits commercial use and modification.