Hugging Face's logo

Datasets:
S3RL4
/
binary-reasoning-v1

Size:
10K<n<100K
Tags:
reinforcement-learning
evaluation
reverse-engineering
binaries
elf
pe
License:
Dataset card Files
xet
 Community
1
binary-reasoning-v1 / DATASHEET.md
S3RL4's picture
S3RL4
Public clean-up: measured public outcomes only; private families shared privately; licence clarity
e67a78d verified
preview code
|
Raw History Blame Contribute Delete
11.1 kB

binary-reasoning-v1 — Datasheet

Product: verified reverse-engineering tasks (dataset + RL/eval environment) over real binaries. Seller: Samuel Anttila / RE-Lab (Rovaniemi, FI). Contact: samueljanttila@gmail.com License: (c) Samuel Anttila / RE-Lab. Free for evaluation & research; commercial license on purchase.

Composition (10,251 items, unique IDs, 2026-10-29)

family tier grading count
func_id_exact — identify internal/stripped function from scrubbed disassembly hard/med deterministic symbol match (alias-tolerant) 9,073
func_id_mc — multiple-choice function identification medium deterministic option match 1,036
func_summary — summarize scrubbed function (purpose + identifying detail) med/hard rubric LLM judge 53
crossarch_equivalence — same function across two ISAs? (+distractors) medium deterministic stance+symbol 25
agentic_flag_hunt — find argv[1] that makes a shipped binary print OK (sandbox-executable) practice exact flag / program stdout + exit-code probe 29
library_ident — SBOM-style: name+version of a production library from artifacts easy/med exact substring (name+version) 14
arch_semantics_qa / isa_semantics_qa — instruction-sequence semantics medium rubric judge 14
pe_import_reasoning — infer program behavior from imports+strings medium rubric judge 4
pe_structure_qa — PE header/section facts easy rubric judge 3

Architectures: x86-64 (Linux ELF, Windows PE), i686, AArch64, ARM32, RISC-V, MIPS(o32/n64), PPC64LE. Sources: Ubuntu production binaries (649 distinct third-party file names referenced in shipped provenance; system libs incl. libcrypto, zstd, sqlite3, libxml2, zlib, libpng, plus Mesa DRI drivers, QEMU device models, systemd internals, Wireshark/PulseAudio/gprofng/rsyslog plugins), Debian production cross-toolchain glibc (AArch64/RISC-V/MIPS64/PPC64LE), Android/Termux production aarch64 libs (libsqlite3 3.54.1, libpng 1.6.59, libarchive 3.8.10, libssh 0.12.2, libcurl 8.23.0, OpenSSL 3.6.5, ncurses 6.5, expat 2.9.0 — platform-tagged in provenance), curated PE samples, cross-compiled targets (honestly labeled).

Automated re-verification through our in-house analysis pipeline (same-day, sampling + full-set)

  • 9,865/9,865 checkable items: answer FUNC verified at exact address & size via readelf.
  • 318/318 sampled disassembly excerpts byte-exact vs objdump; zero unscrubbed names.
  • MC decoy rules hold on all MC items (verified at audit; hardening only removed items, never added); exactly-one-correct-option verified on all (1 defect found & removed).
  • Post-audit hardening: 238 items given full alias acceptance (multiple valid FUNC names at one address), 130 same-size twin items filtered out (answer not uniquely inferable), 18 unresolvable-provenance items dropped. The audit's findings were fixed, not hidden.

Construction & verification

  • Produced by static analysis through our in-house analysis pipeline (Ghidra, binutils objdump/readelf/nm). No target binaries executed for answers.
  • Scrubbing: every symbol equal to the answer or a callee replaced with <...>; automated answer-leak scan through our in-house analysis pipeline: 0/10,251 hits.
  • Ground truth derived from artifact itself (symbol tables, header fields, decompiler output).
  • Auto-grader included (grading.py): deterministic graders for exact/MC/crossarch/flag/library families (exact-match is alias-tolerant); rubric judge for open-ended. Test harness test_grading.py: 10,172 deterministic items scored 1.0 on gold answers, 0.0 on abstention; 79 routed to judge.
  • Full agentic benchmark measured (EVAL_RESULTS-passk.md, 2026-10-29): Claude Opus 5.5 solved 29/29 envs, 145/145 attempts with tools (median 44 s) — every answer byte-exact vs our gold: oracle confirmation that the shipped binary accepts those answers and verifies our grading chain; not a proof that each answer is unique. The public tier is practice/warm-up level for tool-using frontier models. Private task families are calibrated across several models on fresh, never-published instances; results are shared privately with prospective buyers. A pilot can include buyer-run per-model pass-rate calibration with our harness, graders, and analysis report. Single-shot no-tools-requested ID reference stays 7/12 small-n, and a fixed 20-item agentic hard spot check found the sampled command-output/static-analysis families practice-level for Opus with tools (40/40 attempts; non-isolated host run, upper bound; EVAL_RESULTS-frontier.md §C).
  • Agentic challenges: round 1 (8) + round 2 (6, stripped) + round 3 (6: self-decrypting .text via .init_array, ptrace/TracerPid/timing anti-instrumentation, chained-seed stages, masked-blob hash + explicit_bzero, control-flow flattening with decoy states, encrypted-bytecode VM with permuted opcodes) + round 4 (5: two-stage ARX+FNV memory-only chain, MIPS-BE branch-delay-slot trap, Windows PE TLS-callback packer, AArch64 computed-blr chain with PAC noise, RISC-V PC-relative jump-table jalr) + round 5 (4: x86-64 static-pie relocation-table/GOT-indirect hunt, ARM32 Thumb/ARM interworking trap, PPC64LE condition-register juggling gate, PE PEB-walk API hiding with export-name hashing) — the set covers multiple distinct RE techniques; every binary verified OK-on-flag / NO-on-mutation (24 single-char mutations each, zero false OKs); exit status = first failing position enables black-box probing; strings shows no flag fragments.
  • Provenance per item: binary name and sha256_16 on every item; function address/size and architecture where applicable.

Intended use

RL training (GRPO-style verifiable rewards), post-training eval suites, expert-model benchmarks, SFT for code-understanding models. Runnable as-is on Prime Intellect Environments Hub via the bundled verifiers v1 taskset (environment/), or standalone JSONL. Harbor-native: the 29 agentic envs are Harbor task folders (verified: each has task.toml + instruction.md + Dockerfile + tests/test.sh + solution scripts and solution explanations) — runnable with harbor run, and importable for free into HUD ("HUD reads Harbor task folders as they are", docs.hud.ai/platform/harbor-import). Pilot zip: https://huggingface.co/datasets/S3RL4/binary-reasoning-v1/resolve/main/harbor_pilot_29.zip

Positioning vs existing work (surveyed 2026-10-28)

neighbor what it is how binary-reasoning-v1 differs
CrackMeBench (arXiv:2605.10597, 5/2026) educational binary-reasoning agent tasks with executable oracles validates the genre; our set uses production binaries (Ubuntu/Debian glibc/Termux-Android), 10,251 items, and 10 families including cross-arch and SBOM-style identification
knewstimek/crackme-agent-bench (GitHub) small Windows PE reverse-engineering benchmark with MCP debugger tools; analyzed statically through our in-house analysis pipeline: v1 fully static-readable; v2/v3 = TLS-callback packer, .text encrypted at rest, API hidden via PEB-walk, memory-only flag (hash of encoded blob, wiped) different scope: no input channel or oracle; ours adds argv + exit-status probing. Their packing/anti-debug techniques inform our roadmap below
CyberGym (sunblaze-ucb, HF dl≈369k) crash/vuln-finding RL at scale demand proof for binary-security data; different task (dynamic crashes vs static RE reasoning)
cybench-harbor (HF dl≈1.5k) Cybench CTFs in Harbor format precedent that security tasks ship via Harbor — our 29 agentic tasks use the same packaging
AssemblyBench (arXiv:2604.03750 / HF) robotics mechanical assembly — name collision only unrelated
atul10 RE code datasets, ollvm-decompilation (HF dl<400) source-level or decompiler-output pairs no executable oracles, no verification bar, single arch

Conclusion: our niche is verified, scrubbed, oracle-graded production-binary RE plus the agentic env set, not "no such data exists anywhere." Academic function-name-prediction datasets exist at scale (e.g. AGNOMIN arXiv:2509.25514: 396k functions / 3 architectures); our differentiator is verification, provenance, deterministic grading, and the agentic environments. Extension capability demonstrated across two rounds: 9 extension envs built+container-verified spanning x86-64 (chain+FNV; static-pie relocation-table hunt), MIPS BE (delay slots), AArch64 (PAC + computed blr), RISC-V (PC-relative jump table), ARM32 (Thumb/ARM interworking), PPC64LE (condition-register gate) and Windows PE (TLS-callback packing; PEB-walk API hiding) — harbor_round4/ + harbor_round5/, per-env evidence dossiers, demo zips on the HF repo; 31 more designs specified in AGENTIC-ROADMAP.md.

Known limits

  • 79 items route to an LLM judge (53 func_summary, 14 arch/isa-semantics QA, 7 PE reasoning, 5 inferred-stripped func_id); the other 10,172 grade deterministically. Judge model configurable.
  • Some stripped-binary answers are marked inferred:true; graders accept role-correct names.
  • Address coordinate systems vary (PIE load base noted in provenance where relevant).

Files

  • environment/ — verifiers v1 package (taskset.py, grading.py, items.jsonl, bins/ + src/, README.md, pyproject.toml); import-tested against real verifiers v1: all 10,251 tasks load from the full-pack copy inside binary-reasoning-v1.zip; the loose working-tree environment/src/binary_reasoning_v1/items.jsonl is the 100-item open sample
  • verification_kit/ — sample-level public checks (gold.py, do_nothing.py, random_wrong.py, leak_check.py, grader_tamper.py) plus internal hard-family reproduction scripts under solvers/; see verification_kit/README.md for limitations.
  • VERIFICATION-DOSSIER-PUBLIC.md — public summary of verified proof and known caveats; full buyer-ready dossier is private until published.
  • harbor/ — the 29 agentic challenges are packaged in Harbor task format (harbor/ + harbor_round4/ + harbor_round5/) (instruction.md, task.toml, Dockerfile, solution/solve.sh, solution/EXPLANATION.md, tests/test.sh → reward.txt); built and scored in real Docker containers: 29/29 reference solve→reward 1; wrong, tampered-binary, and hard-mode controls →reward 0 (verify_pilot29_docker.sh; RESULT envs_ok=29 envs_bad=0)
  • part_relf/repe/xarch/bulk/agentic/agentic2/isa/glibc/android.jsonl — source batches
  • harbor_round4/+part_agentic4.jsonl+round4/src, harbor_round5/+part_agentic5.jsonl+round5/src — rounds 4–5 proof envs (9 total: x86-64×2/MIPS/AArch64/RISC-V/ARM32/PPC64LE/PE×2), merged into the canonical pack 2026-10-29 (count 10,251, docs re-cut); per-env evidence: harbor_round4/EVIDENCE.md, round5/EVIDENCE-round5.md
  • sample_free.jsonl — 9-item free sample (one per family) for listings
  • test_grading.py, leak_audit.py — verification harnesses