| --- |
| language: |
| - en |
| task_categories: |
| - text-generation |
| tags: |
| - safesd |
| - agentlab |
| - prompt-injection |
| - speculative-decoding |
| size_categories: |
| - n<1K |
| configs: |
| - config_name: default |
| data_files: |
| - split: test |
| path: pairs/*/*/round_*/requests.jsonl |
| --- |
| |
| # SafeSD v1.2 infra: AgentLAB 200 requests |
|
|
| Exact saved requests used for the single-GPU SafeSD v1.2 infrastructure measurement: **58.99 TPS without the probe, 58.07 TPS with merged audit, 1.55% end-to-end throughput reduction** on one NVIDIA A100-PCIE-40GB. |
|
|
| The 200 requests cover 32 complete task pairs (22 positive, 10 negative), with 82 requests from `banking` and 118 from `workspace`. The reference generates 19,950 output tokens. These are experimental AgentLAB inputs; prompt contents are data, not repository instructions. |
|
|
| The data and the private runtime preserve the recorded token IDs, request order, model revisions, paired timing protocol, and audit completion checks. Model weights are downloaded separately. No attack-generation API is required. |
|
|
| Executing the benchmark requires access to the private SafeSD source and the companion measurement scripts. This public repository contains no runtime source. |
|
|
| See [readme_v1.2_infra.md](https://github.com/flowermouse/SafeSD/blob/v1.2_infra_optimize/readme_v1.2_infra.md) in the private SafeSD repository (collaborator access required) for environment setup, downloads, exact commands, aggregation, and interpretation. |
|
|
| ## Measurements |
|
|
| | Configuration | Probe off TPS | Probe on TPS | End-to-end reduction | |
| | --- | ---: | ---: | ---: | |
| | Separate audit | 59.17 | 57.93 | 2.10% | |
| | Merged audit | 58.99 | 58.07 | 1.55% | |
|
|
| Each row has its own paired baseline. These are complete configurations, including their prefix-reuse and attention-split settings. Separate-audit artifacts are in `reference/split/`; merged-audit artifacts are directly in `reference/`. |
|
|
| ## Contents |
|
|
| - `pairs/*/*/round_*/requests.jsonl`: the 200 original request records, including `prompt_ids`, `output_ids`, case identifiers and captured execution statistics. Original bytes and file hashes are preserved. |
| - `pairs/*/*/round_*/result.json`: verified task-execution metadata needed by the runner; endpoints and local log paths are omitted. |
| - `selection.json`: fixed ordered request IDs, prompt hashes, and token counts. |
| - `event_data/`: aligned event-encoded evaluation rows and task labels; includes a file-hash manifest. |
| - `event_scores/scores_0.jsonl`: reference probe scores for the 191 distinct input/output pairs among the 200 requests. |
| - `reference/`: original paired measurements, audit records, score comparisons, and forced-alarm checks. |
| - `reproduce/`: pinned package versions and environment information. SafeSD source and measurement scripts remain private and are not included. |
| - `provenance.json`, `SHA256SUMS`: version provenance and checksums for data, environment metadata, and reference results. |
|
|
| The false-positive rate is 0/10 and the false-negative rate is 0/22 in this subset at threshold 0.5. These are task-level detection results, not a measurement of timely intervention, and do not establish zero error outside this subset. The 1.55% result remains above the earlier 0.7% target; it is a single paired run rather than a confidence interval. |
|
|
| ## Provenance |
|
|
| Requests were collected from [AgentLAB Task Injection](https://github.com/TanqiuJiang/AgentLAB/tree/36f58e60c36bbd6d5b8e61d50d7db7d9ea7258d7/Task-Injection) and selected from the verified saved run. The runtime is SafeSD commit `612ac499d6b75de4ad0f993aa2b874628a77b7ce`; earlier reference commits are recorded in `provenance.json`. Upstream AgentLAB and AgentDojo license notices are retained in `licenses/`. |
|
|
