README.md

---
pretty_name: RLforDecomp
license: other
license_name: upstream-open-source-licenses
task_categories:
- text-generation
tags:
- decompilation
- reverse-engineering
- binary-analysis
- c
- code
size_categories:
- 100K<n<1M
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: val
    path: data/val-*
  - split: test
    path: data/test-*
  - split: decompiler_test
    path: data/decompiler_test-*
dataset_info:
  config_name: default
  features:
  - name: input
    dtype: string
  - name: output
    dtype: string
  - name: function_name
    dtype: string
  - name: repo_name
    dtype: string
  - name: category
    dtype: string
  - name: sample_id
    dtype: string
  - name: decompiler
    dtype: string
  - name: opt_level
    dtype: string
  splits:
  - name: train
    num_examples: 643767
  - name: val
    num_examples: 5954
  - name: test
    num_examples: 10838
  - name: decompiler_test
    num_examples: 48664
---

# RLforDecomp

Function-level pairs for **decompilation refinement**: `input` is a C function as produced by a
decompiler, `output` is the original source function it came from. The task is to reconstruct the
source function from the decompiler output. Quality is measured by the graph edit distance (GED)
between the control-flow graphs of the prediction and of the source function, as in the
[SAILR](https://wilgibbs.com/papers/sailr_usenix24.pdf) paper (GED = 0: identical CFG).
All binaries are x86-64, compiled at `-O0`.

```python
from datasets import load_dataset
ds = load_dataset("Divij/RLforDecomp")   # DatasetDict: train / val / test / decompiler_test
```

## Splits

| split | rows | projects | decompiler(s) | intended use |
| --- | ---: | ---: | --- | --- |
| `train` | 643,767 | 1,972 Gentoo packages | IDA Pro 9.2 | training |
| `val` | 5,954 | 30 Gentoo packages | IDA Pro 9.2 | tuning: checkpoint selection, thresholds, sweeps |
| `test` | 10,838 | 30 Gentoo packages | IDA Pro 9.2 | evaluation on unseen repositories, same decompiler as training |
| `decompiler_test` | 48,664 | 23 SAILR (Debian) projects | IDA (older release), Ghidra, angr-SAILR, angr-Phoenix, angr-DREAM | evaluation on unseen repositories **and** unseen decompilers |

`decompiler_test` rows per decompiler: `ghidra` 11,681, `angr_dream` 11,486, `ida` 11,269,
`angr_sailr` 7,410, `angr_phoenix` 6,818.

Splits are disjoint at the level of the **version-stripped package name** (every version of a project
lands in the same split), and no decompiled input occurs in more than one split. Treat `test` and
`decompiler_test` as held-out: evaluate a frozen system on them once and tune on `val` instead.

## Fields

| field | meaning |
| --- | --- |
| `input` | decompiled function (model input) |
| `output` | original source function (target) |
| `function_name` | function name used to pair the two sides |
| `repo_name` | Gentoo package (with version) or SAILR project |
| `category` | Gentoo category, or `sailr` for `decompiler_test` |
| `sample_id` | parent binary (Gentoo) or SAILR file stem; **not** unique per row |
| `decompiler` | `ida` for train/val/test; one of `ida`, `ghidra`, `angr_sailr`, `angr_phoenix`, `angr_dream` for `decompiler_test` |
| `opt_level` | always `O0` |

A row is uniquely identified by `(sample_id, function_name, decompiler)`.

Length in characters (median / 95th percentile); lengths are heavy-tailed (train inputs: p99 about 11K, max about 454K).

| split | `input` | `output` |
| --- | ---: | ---: |
| `train` | 531 / 4,278 | 447 / 3,094 |
| `val` | 612.5 / 4,696 | 488 / 3,451 |
| `test` | 431 / 4,313 | 449 / 3,233 |
| `decompiler_test` | 625 / 4,694 | 531 / 4,052 |

## How it was built

- **Gentoo (`train`, `val`, `test`).** From the x86-64 Gentoo binaries and source archives released with
  [VarBERT](https://github.com/sefcom/VarBERT) (O0 only; built with debug info and **not stripped**).
  Each binary was matched to its source archive (exact or revision-normalized package name, or a manually
  reviewed override) and its DWARF compilation units to archive members; a binary with any unresolved
  unit was dropped. Every remaining binary was decompiled whole with IDA Pro 9.2 (output normalized with
  the SAILR lexer rules: IDA integer type names mapped to C types, `__fastcall`/`__noreturn`/`__cdecl`
  removed, `::` mapped to `_`), leaving 9,612 binaries (2,107 packages, 76 categories) before pairing and filtering.
- **Pairs.** IDA output is split into functions, top-level functions are extracted from the matched source
  files, and functions are paired by exact, case-sensitive name. IDA auto-generated names (`sub_...`) and
  any name with more than one distinct body on either side are dropped. Comments are removed from both
  sides and both are formatted with clang-format (LLVM style). Exact duplicate (input, output) pairs are removed.
- **SAILR (`decompiler_test`).** All usable O0 output of the [SAILR artifact](https://github.com/mahaloz/sailr-eval):
  23 projects, each source function paired with the output of up to five decompilers. A pair is kept only if
  pyjoern can build a CFG for the named function on both sides. The source side is the artifact's normalized,
  preprocessed source.
- **Contamination guards.** No Gentoo package whose version-independent name matches a SAILR project, no
  Gentoo row whose decompiled input occurs in SAILR, and no exact-input overlap between any two splits.

## Things to know

1. **Unstripped inputs in train/val/test.** Gentoo binaries keep DWARF and symbols, so IDA output for
   `train`/`val`/`test` typically retains real identifiers, types and even label names. `decompiler_test`
   inputs use generic names (`a1`, `param_1`, `struct_0 *a0`, ...), so it measures a substantial distribution shift.
2. **Pairs are matched by name, not proven equivalent.** There is no address- or line-level alignment, and source
   can contain code that the compiler removed (for example `assert` macros). Identical decompiler output can map
   to different sources: in `train`, 5,078 distinct inputs occur with more than one distinct `output` (for example
   compiler/CRT stubs and trivial one-liners).
3. **CFG-buildability was required only for `decompiler_test`.** `train`, `val` and `test` rows were not filtered on it.
4. **Exact deduplication only.** Near-duplicates, forks and vendored copies can remain; leakage checks are exact-match.
5. **Skew.** Large projects dominate: 121 of the 1,972 `train` packages hold half of the rows, two packages hold half
   of `val` and half of `test`, and three projects (openssh-portable, bash, coreutils) hold 49% of `decompiler_test`.
   Read aggregate metrics per repository or with a repository-clustered bootstrap. Because GED is heavy-tailed,
   the exact-match rate (GED = 0) is the more stable summary than the mean.
6. **Verbatim upstream strings.** Functions are copied from public projects and can contain strings that resemble
   credentials (PEM header markers in key parsers; one public Amazon API access-key ID in glyr's `generic_amazon_url`).

## License and provenance

The `output` side is source code from open-source projects (named in `repo_name`) under their own, varying licenses
(for example GPL, LGPL, BSD, MIT); the `input` side is decompiler output for binaries built from that code. This
compilation is provided for research use. You are responsible for complying with the licenses of the upstream projects.

Upstream artifacts:

- VarBERT: K. K. Pal et al., "Len or index or count, anything but v1": Predicting Variable Names in Decompilation
  Output with Transfer Learning. IEEE Symposium on Security and Privacy, 2024.
- SAILR: Z. L. Basque et al., Ahoy SAILR! There is No Need to DREAM of C: A Compiler-Aware Structuring Algorithm
  for Binary Decompilation. USENIX Security Symposium, 2024.