🚧 Status: v0.1 generation in progress (target: Nov 3, 2026). This card describes the dataset that this repo will contain. Generation scripts live in github.com/Hisernberg/tracea.

TraceBench pairs raw runtime telemetry with evidence-grade annotations so agents (and models) can learn and be evaluated on what software actually does at runtime — the dynamic counterpart to static decompilation corpora.

Motivation

Static analysis corpora are abundant (decompiled code, disassembly). Dynamic corpora are not: nobody publishes paired (target, scenario, hook script, syscall/HTTP trace, ground-truth narrative) tuples. TraceBench fills that gap and serves as the training + eval substrate for tracea-explainer.

Planned structure (v0.1 — 200+ tuples; v1.0 — 1,000+)

traces/
  <scenario-id>/
    scenario.yaml        # target, goal, environment, authorizations
    hook.js              # Frida hook script used
    trace.jsonl          # Evidence Record v1 stream (tools: net_*, proc_*, android_*)
    artifacts/           # pcaps, stacktraces, summaries
ground-truth.jsonl       # plain-English behavior narratives per scenario
splits/
  train.jsonl  val.jsonl  test.jsonl   # scenario-level split, no target leakage

Sources

Open-source / intentionally vulnerable apps only: OWASP UnCrackable L1–L4, DIVA, MASTG apps, coreutils, curl, nginx, synthetic test binaries — all authorized targets
Generated with tracea tools (network engine first: pcap → HTTP/TLS/DNS evidence)
Every scenario records explicit authorization + environment metadata

Schema highlights

Field	Meaning
`scenario_id`	stable ID, target + scenario hash
`evidence`	Evidence Record v1 lines (observation, confidence, limitations)
`narrative`	ground-truth plain-English behavior description
`tools`	which tracea tools produced the evidence

Intended uses

Fine-tuning small models to narrate runtime behavior (tracea-explainer)
Benchmarking agent MCP tool-use for dynamic analysis (tracea-eval suite)
Security-education datasets

Out of scope / ethics

No closed-source proprietary targets, no DRM circumvention, no malware. All captures redacted (high-entropy secrets stripped by tracea's redaction engine).

Citation

@dataset{tracebench2026,
  title  = {TraceBench: Annotated Runtime Traces for Dynamic-Analysis Agents},
  author = {Nur, Nabid},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Nabidnur/tracea-bench}
}

TraceBench (tracea-bench) — annotated runtime traces for dynamic-analysis agents