COIN v2026-07

COde-understanding through INput-space mapping — a benchmark that measures an agent's ability to translate a natural-language reachability goal ("reach line X of file Y") into an input that exercises that line at runtime.

This release pins 70 verified targets drawn from 7 OSS-Fuzz projects. Every target is paired with a runtime docker image whose digest is captured below — a snapshot is content-addressed end-to-end, so reruns six months from now hit the same binaries.

Quick start

from datasets import load_dataset

ds = load_dataset(
    "COIN-Bench/coin",
    revision="v2026-07",
    split="codeql_only",
)
print(ds[0]["target_id"], ds[0]["runtime_image"])

To run the full evaluation pipeline against this snapshot:

pip install coin-bench
coin evaluate \
  --dataset COIN-Bench/coin \
  --revision v2026-07 \
  --split codeql_only \
  --output-dir results/

Splits

Splits are first-match-wins over the 6-bit signal cube (G, S, N, F, L, C) — GCS-covered, seed-reachable, fuzzer-reachable (libfuzzer-noseed), baseline-reachable (libfuzzer), LLM-baseline-reachable, CodeQL-statically-reachable.

| Split | Cell | Rows | |---|---|---|| gcs_reachable | G and not (F or L) | 35 || codeql_only | C and not (G or F or L) | 35 |

Runtime images

One per project. Digest-pinned for reproducibility — the dataset rows reference these digests directly.

| Project | Image | Digest | Size | |---|---|---|---|| cups | localhost:5000/cointest/runtime-cups:v2026-07-test1 | sha256:986901f42872c2d3f3705368e9546e9f792efc957a11541540c4232fb7eaf36d | 2698.4 MB || cyclonedds | localhost:5000/cointest/runtime-cyclonedds:v2026-07-test1 | sha256:ebce3ee4f0439dc4137e4102d12d1b0958882af2e519a441e1203be06da72dfd | 3658.2 MB || hdf5 | localhost:5000/cointest/runtime-hdf5:v2026-07-test1 | sha256:23dac6e302371f1fa5341a1e54d537d41168b125a39a1e902464700daaae91a6 | 5762.2 MB || karchive | localhost:5000/cointest/runtime-karchive:v2026-07-test1 | sha256:4c7ea95f50529f13a86fe490cdcd779b3d1c7e699916bda4212dc780d1082665 | 14784.0 MB || liboqs | localhost:5000/cointest/runtime-liboqs:v2026-07-test1 | sha256:b00578ef6bc10a015619ddcf614cb217e8a1ecae49b3da9eba084a3db59e604a | 4118.4 MB || libraw | localhost:5000/cointest/runtime-libraw:v2026-07-test1 | sha256:faa3f5de1db16c0f6b9968ee796375be742b5d8b30f6659ad6150a347c076133 | 2164.6 MB || rdkit | localhost:5000/cointest/runtime-rdkit:v2026-07-test1 | sha256:3361dd349cc9e92dfe48e664b1c0fe61eeadf62cc08834e8f9ae28187f283320 | 12781.0 MB |

Schema

Column	Type	Meaning
target_id	str	<project>:<harness>:<file>:<line_start>[-<line_end>]
coin_version	str	snapshot tag (e.g. v2026-07)
coin_commit	str	git SHA of the COIN repo at release time
oss_fuzz_commit	str	OSS-Fuzz commit pinned for this snapshot
coverage_date	str	GCS HTML coverage report date
project	str	OSS-Fuzz project name
harness	str	primary reaching harness binary
harness_binaries	str (JSON list)	harness binary basenames in the project
language	str	c / c++ / rust / python / jvm
file	str	canonical /src/<project>/<rel> path
line_start, line_end, line_count	int	target line range
function	str	enclosing function (empty if not extracted)
source_snippet	str	±100 lines around the target
src_commits	str (JSON list)	[{path, rev, url}] for git-cloning
gcs_covered	bool	G signal — line in the GCS coverage report
baseline_reachable	str (JSON list or "null")	F — stage5 libfuzzer baseline (with seed if present)
agent_baseline_reachable	str (JSON list or "null")	L — LLM seed-gen baseline
codeql_reachable	str (JSON list or "null")	C — CodeQL static reachability
cnt	int	observed coverage count from GCS
coverage_keys	str (JSON list)	path synonyms for verify-server matching
target_text	str	one-line human description
prompt_preview	str	TASK.md preview that stage7 sends to the agent
runtime_image	str	<image>@sha256:... — the image to spin up
runtime_digest	str	the digest alone, for content addressing
runtime_normal_path	str	path inside the image: harness binaries (/out/normal)
runtime_coverage_path	str	path inside the image: coverage build (/out/coverage)
runtime_source_path	str	path inside the image: source repos (/src)
split	str	denormalized split name (matches the row's file)

What's automated, what's manual

Step	Type
OSS-Fuzz coverage HTML download	manual (operator)
Per-project build (normal + coverage)	manual (operator)
Baseline fuzzing (stages 5 + 5b)	manual (operator)
Target selection (stage 6)	manual (operator)
Snapshot release (coin release)	manual command, fully reproducible
Evaluation against this snapshot	fully automated for downstream users

The selection pipeline is operator-driven; once a snapshot is published, running it against your agent is one command.

Provenance

• COIN repo: https://github.com/coin-bench/coin @ e99b764ba8d7a546e425666851b81545bc800f63
• Source experiment: 20260520T165657Z-b45032068f
• Created at: 2026-07-04T18:44:29.380721+00:00
• Total runtime image footprint: 44.9 GB

License

MIT — see the COIN repository for full terms. Source code referenced from each row remains under its upstream project's license.

Citation

@misc{coinv202607,
  title  = {COIN: COde-understanding through INput-space mapping (v2026-07)},
  author = {COIN Authors},
  year   = {2026},
  url    = {https://huggingface.co/datasets/COIN-Bench/coin/tree/v2026-07}
}