Assemblage-Rust

127,165 compiled Rust binaries from 77,004 builds of 10,766 permissively licensed GitHub repositories — each paired with DWARF-derived function and line metadata, the source tree it was compiled from, and, for a subset, compiler IR.

Every repository is built across a matrix of codegen backends, build modes and optimization levels, so the same source appears many times under different compilation settings. That is the point: it supports studying how compilation choices affect the resulting binary.

Unprocessed. Raw build outputs, filtered only by license — not cleaned, deduplicated or curated. Produced by Assemblage.

Layout

One uncompressed tar per repository (~150 MB average). Everything inside is already compressed — zstd for binaries and metadata, gzip for IR — so the tar is pure container and you can stream one repository without fetching the rest.

repos/{owner}__{project}.tar        builds: binaries, metadata, IR
sources/{owner}__{project}.tar.gz   the source tree each was built from
sources/manifest.json               source file -> repo URL, commit, license, tar
index.jsonl        one line per build: dir, tar, repo_url, commit, license, sizes
LICENSES.csv       build directory -> repo URL, commit, license

repos/X.tar pairs with sources/X.tar.gz — same slug, no lookup table needed. Each tar expands to one directory per build:

{owner}_{project}_{sha12}-{flag}-{backend}-{mode}/
├── binaries/<name>.zst                ELF binaries, zstd -12
├── metadata/assemblage_meta.json.zst  DWARF functions, RVAs, line mappings
├── ir/<stage>.tar.gz                  LLVM-IR / MIR / HIR / THIR / ASM (subset)
└── export.json                        provenance, license, byte counts

Build mode is in the directory name because Debug, RelWithDebInfo and Release of the same commit are distinct binaries.

sources/ ships the builder's own git clone --recursive snapshot, captured before the build ran — no target/, and .git included, so history and the exact tree hash are recoverable. Source correspondence is a property of the release, not of an upstream that may be deleted or force-pushed later.

Quickstart

# one repository's builds + its source tree
huggingface-cli download changliu8541/assemblage-rust \
    repos/tokio-rs__tokio.tar sources/tokio-rs__tokio.tar.gz \
    --repo-type dataset --local-dir .
tar xf repos/tokio-rs__tokio.tar
mkdir -p src && tar xzf sources/tokio-rs__tokio.tar.gz -C src

zstd -d */binaries/mytool.zst -o mytool
zstd -dc */metadata/assemblage_meta.json.zst | jq .

# find builds without downloading anything
jq -r 'select(.license=="MIT License" and .has_ir) | .tar' index.jsonl | sort -u

Build matrix

dimension	values
codegen backend	llvm (50,765) · gcc (16,064) · cranelift (10,175)
build mode	RelWithDebInfo (43,862) · Release (25,716) · Debug (7,426)
optimization	O2 (20,143) · O0 (14,825) · Os (12,474) · O3 (11,512) · Oz (9,677) · O1 (8,373)

Toolchain: nightly-2026-06-15, symbol mangling v0, x86_64-unknown-linux-gnu. 17,174 builds carry IR dumps (llvm-ir, mir, hir, thir, asm), emitted only for the LLVM RelWithDebInfo tiers and scoped to repository crates.

assemblage_meta.json records, per binary, every function recovered from DWARF: function_name (mangled), demangled_name, source_file, function_info (RVA ranges), lines (line -> RVA), and origin ∈ in_repo | dependency | stdlib | other.

Known limitations

Please read these — they are properties of the data, not bugs to work around.

1. ~26% of builds with binaries have empty or partial function metadata. DWARF extraction runs under a wall-clock budget; over budget, the binary still ships but its Binary_info_list entry is dropped. Check its length first.
2. Lib-only crates produce metadata but no binaries — 17,891 builds (23%) are metadata-only.
3. Release binaries carry little repository DWARF by design. Repo symbols survive in .symtab; most DWARF present belongs to the precompiled stdlib. Filter on origin.
4. The gcc (cg_gcc) backend is a name/address corpus. Mangling and symbol tables are sound, but repo-level source_file and line info is largely absent and stdlib paths are recorded relative.
5. origin: in_repo is under-assigned — paths resolve against the clone dir at extraction time, so workspace-relative paths can fall through to other. Treat it as a lower bound.
6. IR builds were compiled with --emit, which repartitions codegen units. .text differs from a non-IR build of the same commit; symbols, .rodata, .data and .eh_frame are byte-identical. DWARF and RVAs come from the published binary and are self-consistent.
7. Binaries over 1 GiB uncompressed are excluded — 317 of them, dominated by cranelift debug info. The build, its metadata and IR are still present.

Licensing and attribution

Every repository carries an OSI-recognized permissive license, verified by database join against the scraper's license records. Copyleft (GPL/AGPL/LGPL/MPL/EPL/EUPL) and unidentified-license repositories are excluded.

license	builds
MIT	55,324
Apache-2.0	18,548
Unlicense	926
BSD-3-Clause	902
BSD-2-Clause	348
CC0-1.0	302
ISC	190
0BSD	186
WTFPL	133
Blue Oak 1.0.0	62
BSL-1.0	45
MIT-0	28
Zlib	6
Artistic-2.0	4

LICENSES.csv maps every build directory to its upstream URL, commit and license; the same fields appear in each export.json and in sources/manifest.json. MIT, BSD and Apache-2.0 require that copyright and license notices accompany redistribution — because sources/ ships each tree verbatim, those notices travel with the data rather than only being pointed at. Shipped trees include .git, so they also carry commit history and its author names and emails, as published by those repositories on GitHub.

If you are an author of an included repository and want it removed, open a discussion on this dataset and it will be taken out.

Citation

@inproceedings{liu2024assemblage,
  title={Assemblage: Automatic Binary Dataset Construction for Machine Learning},
  author={Liu, Chang and Chen, Rebecca Saul and Sun, Yihao and Zhang, Yifan and
          Xing, Gaoyang and Wu, Haoran and Zhu, Zhengzhang and others},
  booktitle={NeurIPS Datasets and Benchmarks},
  year={2024}
}