sriram1983007/sra-bench-results

SRA Stablecoin Risk Bench: Published Results

Raw predictions and scores for models evaluated on the held-out test split (2,000 cases) of sra-stablecoin-risk-bench.

Every model output is published here, so any number below can be checked by re-scoring the predictions with the benchmark's own score.py. No trust in the reported figures is required.

Results

Metric	Base Qwen3-4B-Instruct-2507	SRA-RiskGate-4B
SRA composite score ↑	0.421	0.913
Payments: decision accuracy ↑	59.8%	99.5%
Payments: unsafe approvals ↓	34.6%	0.47%
Payments: over-blocking ↓	5.2%	0.0%
Payments: injected payments approved ↓	52.2%	5.8%
Payments: reason-code (flag) F1 ↑	0.105	0.994
Disputes: score ↑	0.400	0.832
Disputes: outcome accuracy ↑	0.0%	77.8%
Disputes: impossible remedies ↓	n/a †	0.38%
Disputes: wrongful refunds ↓	n/a †	2.2%
Disputes: refund details fully correct ↑	n/a †	64.3%
Schema-valid output (payments / disputes) ↑	99.8% / 0%	100% / 100%

† The base model returns dispute JSON with the right field names 98% of the time, but uses outcomes, mechanisms and flag formats outside the allowed vocabulary (e.g. "outcome": "merchant_wins"), so none of its dispute verdicts are valid and its dispute safety rates are not meaningful.

The full metric breakdown for each run (per-class F1, accuracy by payment format and perturbation, injection statistics and more) is in its scores.json.

Files

Each evaluated model has its own folder:

File	Contents
predictions.jsonl	One line per test case: {"id": "test-000123", "prediction": "<raw model output>"}
test.jsonl	The gold test rows used for scoring (a copy of the benchmark's test split)
scores.json	Full output of score.py for these predictions
run_config.json	Model, backend, decoding settings, precision, timing and token usage

run_eval.py (at the top level) is the script that produced every run.

Re-score a published run

pip install -U huggingface_hub
hf download sriram1983007/sra-bench-results --repo-type dataset --local-dir sra-bench-results
hf download sriram1983007/sra-stablecoin-risk-bench score.py --repo-type dataset --local-dir .

python score.py --gold sra-bench-results/sra-riskgate-4b/test.jsonl \
                --pred sra-bench-results/sra-riskgate-4b/predictions.jsonl

Reproduce a run from scratch

Both runs used the same setup:

• Prompts: the exact training prompts from the benchmark's sft split (task-specific system prompt plus the tagged user message), identical for every model
• Decoding: greedy (temperature 0), up to 512 new tokens
• Hardware and engine: one NVIDIA T4 (float16) with vLLM 0.31, 32 parallel requests

pip install vllm datasets
vllm serve sriram1983007/SRA-RiskGate-4B --served-model-name sra-riskgate-4b \
    --dtype half --max-model-len 12288 --port 8000 &

export VLLM_KEY=none
python run_eval.py --name sra-riskgate-4b --backend openai \
    --base-url http://localhost:8000/v1 --api-key-env VLLM_KEY \
    --model sra-riskgate-4b --workers 32

Determinism: two independent full runs of SRA-RiskGate-4B, on different machines, produced identical scores.

Add your model

Results for other models are welcome. Run run_eval.py on your model (use the same prompts and greedy decoding), then open a discussion on this repo with your predictions.jsonl and run_config.json. Submissions are re-scored with score.py before being added.

Related

• Model: SRA-RiskGate-4B
• Benchmark: sra-stablecoin-risk-bench
• Live demo: sra-riskgate-demo