File size: 2,357 Bytes
afcfd4d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f9db401
afcfd4d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: cc-by-4.0
task_categories:
  - text-generation
  - question-answering
language:
  - en
pretty_name: ShellOps
tags:
  - cli-agents
  - reinforcement-learning
  - filesystem
  - shell
  - verifiable-evaluation
---

# ShellOps

ShellOps is a verifiable dataset suite for CLI agents that interact with filesystem workspaces through executable shell actions. It accompanies the paper [Learning CLI Agents with Structured Action Credit under Selective Observation](https://arxiv.org/abs/2605.08013) and the code release at [Agentic-RL-A3](https://github.com/Hoyant-Su/Agentic-RL-A3).

The dataset focuses on shell-driven information extraction and file editing tasks, where an agent receives a natural language instruction, observes an initial workspace, executes CLI actions, and is scored by verifiable terminal output or post-execution file state.

## Dataset Overview

- **ShellOps**: 1624 standard CLI tasks across training and evaluation splits
- **ShellOps-Pro**: 150 harder out-of-distribution tasks for infer-only evaluation
- **Task axes**: Lookup, Aggregate, Edit, and Mixed
- **Workspace setting**: repository-like file trees with structured data, logs, code, prose, configuration files, and extensionless files
- **Evaluation target**: exact stdout, file-state match, or hybrid output plus file-state objectives

## File Layout

```text
shellops/
  train_src.parquet
  train.parquet
  test.parquet
  assets/

shellops_pro/
  test.parquet
  assets/
```

In `shellops`, `train_src.parquet` is the full training-side corpus claimed in the paper, while `train.parquet` is the sampled subset used for the reported training runs. The `train_src.parquet` and `test.parquet` files together form the 1624-task ShellOps corpus.

`shellops_pro` is an OOD infer-only dataset, so only `test.parquet` is provided. If you use our TEST MODE workflow, create a placeholder `train.parquet` in `shellops_pro/` to keep the file layout consistent. This placeholder parquet is not used during inference.

## Citation

```bibtex
@misc{su2026learningcliagentsstructured,
      title={Learning CLI Agents with Structured Action Credit under Selective Observation}, 
      author={Haoyang Su and Ying Wen},
      year={2026},
      eprint={2605.08013},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2605.08013}, 
}
```