File size: 5,593 Bytes
cb0011e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 | ---
license: mit
language:
- en
tags:
- swe
- dockerless
configs:
- config_name: swe
data_files: "SWE-DockerLess.jsonl"
---
# SWE-DockerLess
A SWE-bench-shaped benchmark that needs **no Docker, no network and no prebuilt
image**. Every row carries its own repo snapshot, so an instance is reproduced by
writing files to a temp directory, `git init`, `git apply` and running pytest.
Rows are byte-for-byte SWE-bench compatible: same field names, same
`FAIL_TO_PASS` / `PASS_TO_PASS` JSON-string encoding, same "resolved" rule. One
extra `dockerless` object carries the snapshot and the test command.
```
SWE-DockerLess.jsonl 4 instances, 62.9 KiB (~16k tokens total)
build_dataset.py rebuilds the jsonl from repos/ + fixed/ + newtests/ and
derives F2P/P2P by actually running pytest
run_harness.py scoring harness (--gold, or --predictions preds.jsonl)
repos/ the four buggy base repos (the source of the snapshots)
fixed/ gold-fix versions of the touched files
newtests/ the tests that become test_patch
tools/validate.py schema, patch-application and budget checks
tools/make_controls.py generates empty / cross / partial negative controls
reports/ harness output for gold and the three controls
```
## Instances
| instance_id | repo | the bug | F2P | P2P | gold patch |
| --- | --- | --- | --- | --- | --- |
| `swedl__miniconf-1` | swedl/miniconf | `${k}` interpolation recurses on cycles (no `CircularReferenceError`) and `$$` never escapes | 9 | 15 | 41 added lines |
| `swedl__ttlcache-1` | swedl/ttlcache | LRU order ignores reads and overwrites; `len()` counts expired rows | 5 | 12 | 12 added lines |
| `swedl__cronlite-1` | swedl/cronlite | `5/15` collapses to `{5}`; day-of-week `7` rejected | 8 | 21 | 11 added lines |
| `swedl__semverlite-1` | swedl/semverlite | prerelease ordering is lexicographic; prereleases satisfy release ranges | 7 | 19 | 23 added lines |
29 FAIL_TO_PASS tests, 67 PASS_TO_PASS tests. Difficulty is SWE-bench-like rather
than toy-like: each fix needs a behaviour change in a real algorithm (recursive
descent with cycle detection, an LRU invariant, cron field expansion, semver
precedence), and the hidden tests pin exact semantics — error message text, the
0-6 weekday scale, the "prerelease must opt in" range rule — that a shallow
patch misses. The `partial` control below scores 5/9, 2/5, 6/8 and 3/7 F2P
without resolving a single instance.
## Row schema
| field | notes |
| --- | --- |
| `instance_id` | `swedl__<repo>-1` |
| `repo` | `swedl/<repo>` |
| `base_commit` | real 40-char sha of the snapshot commit (deterministic) |
| `patch` | gold fix, `git diff` unified diff against `base_commit` |
| `test_patch` | adds test files only, never touches source |
| `problem_statement` | issue-style report; states the contract, never names a test |
| `hints_text` | `""` |
| `created_at`, `version` | `2026-10-30`, `1.0` |
| `environment_setup_commit` | equals `base_commit` |
| `FAIL_TO_PASS`, `PASS_TO_PASS` | JSON-encoded strings, as in SWE-bench |
| `dockerless.files` | `{path: content}` snapshot of the repo at `base_commit` |
| `dockerless.test_cmd` | `python -m pytest -v --tb=no -p no:cacheprovider {tests}` |
| `dockerless.test_files` | the node-id prefixes to run |
## Running it
```bash
python -m pip install pytest # the only dependency
python build_dataset.py # regenerate + re-verify the jsonl
python run_harness.py --gold # sanity check, expect 4/4
python run_harness.py --predictions predictions/partial.jsonl --out reports/partial.json
```
A predictions file is JSONL of `{"instance_id", "model_name_or_path",
"model_patch"}` — the SWE-bench format. `resolved` requires **all** F2P and
**all** P2P to pass.
## Verified results
```
$ python run_harness.py --gold
instance resolved F2P P2P note
swedl__miniconf-1 yes 9/9 15/15 ok
swedl__ttlcache-1 yes 5/5 12/12 ok
swedl__cronlite-1 yes 8/8 21/21 ok
swedl__semverlite-1 yes 7/7 19/19 ok
resolved 4/4 (100.0%) in 10.3s
$ python run_harness.py --predictions predictions/empty.jsonl -> 0/4
$ python run_harness.py --predictions predictions/cross.jsonl -> 0/4
$ python run_harness.py --predictions predictions/partial.jsonl -> 0/4
```
`build_dataset.py` is the ground truth for F2P/P2P: it materialises a clean
checkout, applies `test_patch`, runs pytest, applies `patch`, runs pytest again,
and classifies every test by observed outcome. It refuses to write a row if the
gold patch leaves any failure behind, if F2P is empty, or if a test disappears.
## Budget
Measured for `max_tokens = 2048`, `context = 240000`:
- prompt per instance: ~2.0k–2.6k tokens (problem statement + snapshot)
- gold patch: 11–41 added lines, ~200–700 tokens — comfortably inside 2048
- whole dataset: ~16k tokens, so all four rows fit in one context if you want
## Deliberate deviations from SWE-bench
- **Synthetic repos.** These are four purpose-built Python packages, not real
GitHub projects, so there is no `repo` clone step and no upstream history.
- **Snapshot in the row** instead of a Docker image. `base_commit` is a genuine
commit sha of that snapshot, so `git apply` and `git diff` behave normally.
- **No install step.** Packages are importable from the repo root, so
`dockerless.install` is empty and pytest runs with the repo as rootdir.
- **Python only**, single language, to keep the harness to `python + git + pytest`. |