File size: 4,830 Bytes
3617bce
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: mit
task_categories:
- text-generation
language:
- en
tags:
- agents
- agentic-benchmark
- evaluation
- pinchbench
size_categories:
- n<1K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*.parquet
---

# PinchBench (nearai-bench packaging)

A **flat, self-contained repackaging** of the upstream
[PinchBench](https://pinchbench.com) skill suite (147 tasks,
BENCHMARK_VERSION 2.0.0) for agent-harness consumption. Task content is
**unmodified** — ids, filenames, prompts, rubrics and graders are byte-identical
to upstream, so scores stay comparable to the
[pinchbench.com](https://pinchbench.com) leaderboard.

## Why this exists

Upstream ships tasks as a directory of markdown files plus a separate asset pool
that has to be fetched with a shell script (and Git LFS). That is awkward for an
eval/RL harness that wants to stream tasks across workers. Here, each task is
**one row**: the verbatim markdown in `task_md`, and every asset the task
references as a deterministic `tar.gz` in `assets_tar`. No clone, no LFS, no
fetch script.

```python
from datasets import load_dataset
ds = load_dataset("NEAR-AI/pinchbench", split="train")
```

## Columns

| Column | Type | Notes |
|---|---|---|
| `task_id` | string | Upstream task id, matches the upstream filename stem |
| `name` | string | Human-readable title |
| `category` | string | One of the 11 upstream categories |
| `grading_type` | string | `automated` \| `llm_judge` \| `hybrid` |
| `timeout_seconds` | int64 | Upstream per-task budget |
| `prompt` | string | `## Prompt` section |
| `expected_behavior` | string | `## Expected Behavior` section |
| `automated_checks` | string | `## Automated Checks` — Python `grade(transcript, workspace_path)` |
| `llm_judge_rubric` | string | `## LLM Judge Rubric`, empty when the task has none |
| `grading_criteria` | string | `## Grading Criteria`; upstream judge falls back to this when the rubric is empty |
| `grading_weights` | string (JSON) | `{"automated": w, "llm_judge": w}`; `{}` ⇒ upstream 0.5/0.5 default |
| `multi_session` | bool | Multi-turn dialogue task |
| `sessions` | string (JSON) | Session definitions for multi-session tasks |
| `prerequisites` | string (JSON) | External tooling the task needs, e.g. `["cli:gh"]` |
| `workspace_files` | string (JSON) | Upstream `workspace_files` verbatim — inline `content` entries and `source`/`dest` asset refs |
| `asset_paths` | string (JSON) | Asset paths bundled in `assets_tar` |
| `assets_tar` | string | base64(tar.gz) of this task's assets; `""` when the task has none |
| `task_md` | string | **The verbatim upstream markdown file**, frontmatter included |

## Reconstructing an on-disk task

`task_md` + `assets_tar` reproduce the upstream layout byte-for-byte, which is
what keeps grading identical:

```python
import base64, io, json, tarfile, pathlib

def materialize(row, out_dir):
    out = pathlib.Path(out_dir); out.mkdir(parents=True, exist_ok=True)
    (out / f"{row['task_id']}.md").write_text(row["task_md"])
    if row["assets_tar"]:
        blob = base64.b64decode(row["assets_tar"])
        with tarfile.open(fileobj=io.BytesIO(blob), mode="r:gz") as tar:
            tar.extractall(out / "assets")
    return out
```

Then seed a task workspace from `workspace_files`: entries with `path` +
`content` are written inline; entries with `source` + `dest` are copied from the
extracted `assets/` tree.

## Grading

Upstream's three modes, unchanged:

- **automated** — run the Python `grade()` in `automated_checks` over the agent
  transcript and workspace; task score is the arithmetic mean of the returned
  criterion scores.
- **llm_judge** — score against `llm_judge_rubric` (or `grading_criteria` when
  the rubric is absent); the judge's `total` is the task score.
- **hybrid** — weighted combination using `grading_weights`, defaulting to
  0.5/0.5.

A reference implementation of all three (a faithful port of upstream
`lib_grading.py`) lives in
[`nearai/benchmarks`](https://github.com/nearai/benchmarks) at
`src/adapters/pinchbench.rs`.

## Caveats

- 4 tasks (`gws_*`, `gh_issue_triage`) declare `prerequisites` — npm
  `@juppytt/fws`, the `gh`/`gws` CLIs — that no loader installs for you. They
  score near zero without that tooling (and without a mock backend for the
  live services they drive).
- Assets are scoped to what the 147 upstream tasks actually reference.

## Provenance & license

- **Upstream**: <https://github.com/pinchbench/skill> — MIT, by
  [Kilo Code](https://kilo.ai) (@olearycrew, @arpitg1991, @DJRHails,
  @evanjacobson, @iJaack).
- **This repackaging**: MIT, same terms. Task content unmodified; only the
  container format changed.
- **Packaged by**: [NEAR AI](https://near.ai) for
  [nearai-bench](https://github.com/nearai/benchmarks).