macmacmacmac/Sev-behavioral-research-v1

Sev security evidence collection

Sev-4B v0.3.0 research update

The v0.3.0 response-policy checkpoint improves authored policy decisions from 127/175 to 147/175 across 35 held-out source programs. At its calibration-selected alert threshold, it flags 14/140 permitted decisions, a 10% false-alert rate on this panel. 26/28 registered checks pass. The DNS diagnostic remains 21/32, with one repaired answer and one newly incorrect answer. These are bounded research results, not field validation or proof of human-versus-agent classification.

The current 40-entry collection inventory and catalog preserve all previous entries. The new response-policy configuration contains 139 training, 28 calibration, and 35 development programs with explicit authored permissions. It interprets recovered code; it does not establish historical authorization, execution success or actor identity. Training contains two policy questions per program; calibration and development contain five. Keep the frozen groups and partitions.

The DNS configuration is evaluation-only: 8 summaries and 32 questions share one incident-inspired group. Five summarize a public incident report; three are authored, unexecuted controls. No publisher reasoning, actor attribution or verdict enters the model state. These summaries are not raw telemetry or a complete transcript. Do not train, calibrate or choose thresholds on this panel. The combined curriculum and transfer entries publish metadata only; mixed-term replay payloads remain withheld. Source terms are separate. SwarmTraces reuse terms remain unverified, and this release grants no blanket source-content license.

Security logs, publisher-reconstructed artifacts, checked task derivatives and simulated research data for evidence-based defender models. The collection catalog has 40 entries, including sources with downloadable files, sources linked to their publishers, restricted local material and a retired-data record. Each entry states its provenance, label scope, reuse terms and limitations.

Browse the datasets · Collection inventory · Machine-readable catalog · Sev source

Source mirrors live under collection/v2/sources/<source-id>/; the new frozen SwarmTraces suites live under collection/v2/experiments/swarmtraces-v1/. The catalog pins their byte counts and SHA-256 values. It includes the current behavioral corpus, Splunk BOTS and Attack Data, OTRF/Mordor, EVTX, selected public behavior and diagnostic datasets, ExCyTIn, Microsoft GUIDE, current source-specific Sev derivatives, and SwarmTraces. Some mixed curricula publish metadata only while their component reuse terms are reviewed. Source availability does not mean that every file is admitted to model training.

SwarmTraces in Sev format

The new conversion preserves all 189,579 artifacts in the pinned SwarmTraces source. It stores exact text in observations.jsonl.gz, source provenance in annotations.jsonl.gz, and unknown actor, maliciousness, ownership and execution labels in labels.jsonl.gz. The manifest and audit document the conversion and its round-trip checks.

from huggingface_hub import hf_hub_download

manifest = hf_hub_download(
    "macmacmacmac/Sev-behavioral-research-v1",
    "collection/v2/sources/swarmtraces_sev_v1/manifest.json",
    repo_type="dataset",
)

All source timestamps are null. Reconstruction parent links and response categories do not establish observed request/response causality or successful execution. Exact text repeats across roots, so root-only train/test splits are insufficient. The observation envelope groups reconstructed artifacts for storage; it does not claim that those groups are measured runtime sessions. Upstream reuse terms are unverified; no new license is assigned to this source.

Frozen SwarmTraces code-interpretation suites

Suite	Training	Calibration	Development	Published here
Static v1	446	106	176	Non-test records and provenance
Manual v1	0	0	40	Development records and annotation receipts
Combined curriculum v1	6,438	1,892	1,793	Metadata only
Combined transfer v1	0	0	177	Metadata only

The static labels come from conservative program-analysis rules. The manual panel contains 83 independently authored answers and is evaluation-only. It was separated from automatic selection before model scoring; it is a development screen, not an untouched final test. These questions concern what the supplied recovered code means, not whether it ran or who controlled it. Counts for the mixed suites describe their withheld source partitions. Their mixed-source replay payloads are not mirrored. All four original manifests remain byte-identical, and no test partition is included in this update. The new suites retain the source's unresolved reuse terms.

Original frozen behavioral release

The original files remain unchanged under behavioral-research-v1/:

Partition	Records
Training	2,334
Calibration	482
Development	1,265
Empty test file	0

These are wholly generated records from authored policies in a stateful simulator: 2,094 behavioral training windows and 240 upstream authored rule records. They do not contain captured people or executed LLM agents. Keep counterpart policies, windows and observation-loss views in their original group and partition. Original behavioral records use CC BY 4.0; the authored Kev rules retain Apache-2.0. LICENSE.md applies to that original release, not to every added source in the collection.

Evidence and label boundaries

In supervised Sev records, state is the model-visible evidence, questions contains targets, and _meta contains audit annotations. Never concatenate the targets, source labels or audit metadata into the model input. The native simulator ledger also contains latent truth and is intended for audit and reprojection.

Public source labels have different meanings. A security attack label does not establish agent origin; a recovered program does not prove it ran; simulated controller policies do not establish real-world actor identity. Private device captures and retired generated payloads are excluded from the new mirrors. No locked Sev final test or model weights are included in this collection update. Follow the license and attribution evidence for each source.