File size: 5,725 Bytes
4f713aa
fec0c33
 
 
 
 
 
 
 
 
 
 
 
 
 
4f713aa
 
 
 
 
 
 
 
fec0c33
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
---
license: mit
task_categories:
- text-generation
language:
- en
tags:
- code
- agent
- benchmark
- tool-use
- swe
pretty_name: SACB (Simple Agent Coding Benchmark)
size_categories:
- n<1K
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*
  - split: extended
    path: data/extended-*
---

# SACB — Simple Agent Coding Benchmark

A multi-turn agentic coding benchmark that needs **no Docker**. The model is
given a small but real repository, a deliberately narrow tool set, and a bug
report. It drives its own conversation until it declares itself finished, and is
then graded by running a test suite it was never allowed to see.

Built for [`llama-eval`](https://github.com/ggml-org/llama.cpp) (`--dataset agentic`),
but the records are self-contained and usable by any harness.

## Why it exists

Most agentic coding benchmarks need a container per task, because each one drags
in a different framework and toolchain. SACB deliberately restricts itself to
two languages with lightweight native sandboxes and a tiny dependency set, so a
run needs a venv and, for the TypeScript half, node — nothing more.

## Design

Three constraints, all deliberate:

**No shell, and no way to run the tests.** The only feedback channel is a
`lint` tool. A syntax or type error can be driven out mechanically; a logic
error has to be reasoned about. A harness that lets an agent iterate against the
test suite measures something closer to search than to understanding.

**Two edit tools**, one line-addressed and one content-addressed. Which one a
model reaches for, and whether it keeps line numbers straight after its own
earlier edit, is itself a signal.

**Tests never touch the working tree.** They are held outside it and copied into
a throwaway copy only after the agent stops, so they can be neither read nor
edited.

The expected tool set is `list_files`, `read_file` (with line-range narrowing),
`search`, `edit_lines`, `edit_replace`, `write_file`, `lint`, `finish`.

## Splits

| split | tasks | contents |
|---|---|---|
| `test` | 60 | the default benchmark |
| `extended` | 129 | everything validated, including the easy `ledger` tier |

The `test` split is a deliberate selection, not a sample. Difficulty here is a
property of *composition*: it is 10 single-defect tasks plus 50 compound tasks,
which is what places a leading small model in the intended band.

## Repositories

Every task is an overlay on one of three hand-written base repositories. The
base is the *correct* code, so the reference fix cannot fail to work.

| repo | language | size | role |
|---|---|---|---|
| `flow` | Python | 989 lines, 13 modules | workflow scheduler |
| `router` | TypeScript | 578 lines, 12 modules | HTTP router |
| `ledger` | Python | 417 lines, 6 modules | event-sourced inventory (easy tier, `extended` only) |

## Fields

| field | meaning |
|---|---|
| `task_id` | unique id |
| `repo`, `lang`, `category`, `difficulty` | metadata |
| `instruction` | the bug report shown to the model |
| `files` | JSON object: path → contents. The starting tree |
| `tests` | JSON object: hidden tests, **never** placed in the agent's tree |
| `gold` | JSON object: the reference fix, for validation |
| `fail_to_pass` | tests that must go from failing to passing |
| `pass_to_pass` | tests that must stay passing |
| `n_tests` | total tests |

`files`, `tests` and `gold` are JSON strings; parse with `json.loads`.

## Scoring

A task is **resolved** when every `fail_to_pass` test passes *and* no
`pass_to_pass` test broke. Report the fraction of `fail_to_pass` alongside:
many tasks carry several independent defects, and that fraction separates
"fixed two of the three faults" from "changed nothing".

`fail_to_pass` and `pass_to_pass` are **derived**, never hand-written — the
tests are run once against the defective tree and once against the reference,
and the sets fall out of the difference. A task whose defect no test exercises,
or whose reference fix does not itself pass, is rejected rather than shipped.

## Calibration

Measured against **Qwen3.5-4B** (Q6_K), 19 episodes on the hard repositories:

| tier | resolved |
|---|---|
| single defect | 3/5 |
| compound (2–4 defects) | 1/14 |

The `test` split projects **~16%** on those rates. Typical episode: 4–12 turns,
16–134k cumulative tokens (median ~40k), 4–11k peak context.

### What actually controls difficulty

Measurement contradicted the obvious guesses, so they are worth recording:

* **Repository size dominates defect count.** A *single*-defect task on `flow`
  went unresolved while a *three*-defect compound on `ledger` resolved
  completely. What makes these tasks hard is orienting in a repository too large
  to read at once. To make the benchmark harder, add a larger repository — not
  more defects per task.
* **Sub-defect success is correlated, not independent.** Two-defect compounds
  resolve near 50%, not the 25% that multiplying probabilities predicts: the
  expensive part is orienting in the code, and that cost is paid once and shared
  across every defect in the same task.
* **Instruction vagueness barely matters.** Rewriting reports from diagnosis to
  bare symptom moved the number far less than expected.

### A harness note worth heeding

If your harness ends an episode as soon as the model emits no tool call, you
will measure a **fake 0%**. Models routinely narrate their analysis in prose
mid-task; treating that as "done" ends the episode with no edits. Send a neutral
nudge and let it continue — roughly one episode in four exits that way.

## License

MIT. All repositories, defects and tests are original work written for this
benchmark; no upstream code is redistributed.