Assemblage-Rust

An unprocessed Rust dataset crawled from GitHub: raw build outputs, filtered only by license. Not cleaned, deduplicated, or curated.

A corpus of 18,450 compiled Rust binaries built from 3,034 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 intermediate representations.

Produced by Assemblage, a distributed binary-corpus generator. 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 — which is the point: it supports studying how compilation choices affect the resulting binary.

Layout

The corpus ships as one uncompressed tar per upstream repository (3,033 of them, ~105 MB on average). Everything inside is already compressed — zstd for binaries and metadata, gzip for IR — so the tar is pure container, and you can stream a single 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

Builds and source are keyed on the same slug, so repos/X.tar pairs with sources/X.tar.gz — no lookup table needed for the common case.

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

The build mode is part of the directory name because Debug, RelWithDebInfo and Release of the same commit are distinct binaries.

Source

sources/ ships the exact tree each binary was compiled from — 3,033 archives, ~31 GiB, one per repository, at the single commit recorded in index.jsonl. They are the builder's own git clone --recursive snapshot, captured before the build ran, so they contain no target/ directory and no build output. .git is included, so history and the exact tree hash are recoverable.

This matters because the alternative is depending on GitHub: a repository that is deleted, made private or force-pushed after collection would otherwise leave its binaries permanently unaccompanied, with DWARF naming source_file paths nothing could resolve. Source correspondence is a property of the release, not of an upstream that may not survive it.

# fetch and unpack one repository
huggingface-cli download changliu8541/assemblage-rust \
    repos/tokio-rs__tokio.tar --repo-type dataset --local-dir .
tar xf repos/tokio-rs__tokio.tar

# fetch the matching source tree
huggingface-cli download changliu8541/assemblage-rust \
    sources/tokio-rs__tokio.tar.gz --repo-type dataset --local-dir .
mkdir -p src-tokio && tar xzf sources/tokio-rs__tokio.tar.gz -C src-tokio

# read a binary and its metadata
zstd -d */binaries/mytool.zst -o mytool
zstd -dc */metadata/assemblage_meta.json.zst | jq .

# resolve a DWARF source_file against the shipped tree
zstd -dc */metadata/assemblage_meta.json.zst \
  | jq -r '.Binary_info_list[0].functions[] | select(.origin=="in_repo") | .source_file' \
  | sort -u | head

# 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 (12,622) · cranelift (2,986) · gcc (2,842)
build mode	RelWithDebInfo (10,264) · Release (6,145) · Debug (2,041)
optimization	O2 (5,599) · O0 (3,820) · Os (2,573) · O3 (2,453) · Oz (2,198) · O1 (1,807)

Toolchain: nightly-2026-06-15, symbol mangling v0, x86_64-unknown-linux-gnu.

4,278 builds carry IR dumps (llvm-ir, mir, hir, thir, asm), emitted only for the LLVM RelWithDebInfo tiers and scoped to repository crates rather than dependencies.

Metadata

assemblage_meta.json records, per binary, every function recovered from DWARF:

{
  "function_name": "_RINvNtCs...",       // mangled
  "demangled_name": "core::ptr::drop_glue::<...>",
  "source_file": "src/engine.rs",
  "function_info": [{"rva_start": "...", "rva_end": "..."}],
  "lines": [{"line_number": 42, "rva": "...", "length": 10}],
  "origin": "in_repo | dependency | stdlib | other"
}

Known limitations

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

1. ~26% of builds that ship binaries have empty or partial function metadata. DWARF extraction runs under a wall-clock budget; when a binary exceeds it the binary is still published but its Binary_info_list entry is dropped. Check Binary_info_list length before assuming coverage.
2. Lib-only crates produce metadata but no binaries. Roughly a quarter of builds are metadata-only.
3. Release binaries carry little repository DWARF by design. Repository symbols survive in .symtab; most DWARF present belongs to the precompiled standard library. Filter on origin.
4. The gcc (cg_gcc) backend is a name/address corpus. Mangling and symbol tables are sound, but repository-level source_file and line information is largely absent, and standard-library paths are recorded relative.
5. origin: in_repo is under-assigned. Classification resolves paths against the clone directory at extraction time, so workspace-relative paths can fall through to other. Treat in_repo as a lower bound.
6. Builds with IR were compiled with --emit, which repartitions codegen units. .text bytes differ 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.

Licensing and attribution

Every repository in this corpus 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 were excluded from this release.

license	builds
MIT	13,241
Apache-2.0	4,448
BSD-3-Clause	214
Unlicense	177
BSD-2-Clause	91
0BSD	73
ISC	55
CC0-1.0	54
Blue Oak 1.0.0	34
WTFPL	29
BSL-1.0	27
Artistic-2.0	4
Zlib	2
MIT-0	1

LICENSES.csv maps every build directory to its upstream repository URL, commit and license; the same fields appear in each build's 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 repository's tree verbatim, those notices — LICENSE, NOTICE, per-file headers — travel with the data rather than only being pointed at, which is the condition those licenses actually ask for. The recorded URL and commit identify the upstream each one came from.

Note that the shipped trees include .git, so they also carry commit history and its author names and email addresses, 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

If you find this dataset useful, please consider citing our paper: Assemblage: Automatic Binary Dataset Construction for Machine Learning.

@article{liu2024assemblage,
  title={Assemblage: Automatic Binary Dataset Construction for Machine Learning},
  author={Liu, Chang and Saul, Rebecca and Sun, Yihao and Raff, Edward and
          Fuchs, Maya and Southard Pantano, Townsend and Holt, James and
          Micinski, Kristopher},
  journal={arXiv preprint arXiv:2405.03991},
  year={2024}
}