File size: 4,446 Bytes
f9c74c0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: cc-by-4.0
language:
- en
pretty_name: OpenClawBench
size_categories:
- 10K<n<100K
task_categories:
- text-classification
tags:
- llm-agents
- agent-trajectories
- anomaly-detection
- tool-use
- function-calling
- evaluation
configs:
- config_name: trajectories
  default: true
  data_files:
  - split: train
    path: data/train.jsonl
  - split: test
    path: data/test.jsonl
- config_name: human_verified
  data_files:
  - split: human_verified
    path: data/human_verified.jsonl
---

# OpenClawBench

**OpenClawBench: Benchmarking Process-side Anomalies in Real-world Agent Execution Trajectories**
(NeurIPS 2026, Evaluations & Datasets Track)

Code: https://github.com/YangzeLiu/OpenClawBench

OpenClawBench labels *process-side* anomalies in LLM agent trajectories: cases where an agent
commits to an action or a confident answer while the information it needs is missing,
ambiguous, or contradicted, whether or not the task outcome looks correct. Trajectories were
generated by six open-weight backbones running as tool-using agents on BFCL tasks. Each
trajectory carries a binary anomaly label, one of five anomaly subtypes, an onset step, and
cited evidence steps.

## Files

| File | Rows | Content |
|---|---|---|
| `data/train.jsonl` | 23,857 | Trajectories with silver labels (training pool) |
| `data/test.jsonl` | 2,646 | Trajectories with silver labels (held-out test) |
| `data/human_verified.jsonl` | 440 | Human labels for a subset of the trajectories above |
| `TAXONOMY.md` | | Anomaly definition, the five subtypes, decision rules, boundary cases |
| `SCHEMA.md` | | Field-level schema |
| `croissant.json` | | Croissant 1.0 metadata with Responsible AI fields |

```python
from datasets import load_dataset
traj = load_dataset("YangzeLiu/OpenClawBench", "trajectories")       # train / test
human = load_dataset("YangzeLiu/OpenClawBench", "human_verified")    # 440 human labels
```

Join `human_verified` to trajectories on (`source_model`, `trajectory_id`).

## How the labels were produced

**Silver labels.** An LLM judge (DeepSeek-V4-Pro) annotated every trajectory with a
self-contained 5-class prompt (in the code repository). Released rows are the
high-confidence subset, split 9:1 into train and test, stratified by source model.

**Human labels (`human_verified.jsonl`).** Two annotators labeled each trajectory
independently, blind to each other and to the silver labels; disagreements were resolved by
adjudication under written boundary-case rules (`TAXONOMY.md`). Labels are binary plus
subtype. Two subsets:

- `gold300`: 300 test trajectories, 50 per source model.
- `supp140`: 140 trajectories stratified by silver subtype to cover the rare subtypes
  (25 silver-normal; 25 each of `capability_gap_overcommitment`, `weak_evidence_commitment`,
  `write_under_unresolved_ambiguity`; 20 each of `premature_external_write`,
  `error_ignored_escalation`), disjoint from `gold300`. Because of the stratification its
  anomaly rate is not representative of the corpus.

## Source models

DeepSeek-R1, GLM-4.5-Air, GPT-OSS-20B, Qwen3.5-9B, Qwen3.6-27B, Qwen3.6-35B-A3B,
roughly balanced across the corpus.

## Intended use and limitations

Intended for training and evaluating trajectory-level anomaly detectors, subtype
classification, onset localization, and studying LLM-judge reliability against human labels.

- Silver labels come from a single LLM judge and contain errors; use `human_verified` for
  evaluation claims.
- Only open-weight backbones on BFCL function-calling tasks, English only; behavior of
  proprietary agents or open-ended tasks is not covered.
- Subtype boundaries (notably `weak_evidence_commitment` vs. `capability_gap_overcommitment`)
  are inherently fuzzy; see `TAXONOMY.md`.
- Trajectories are model outputs on public benchmark prompts and contain no personal data.
  API keys and access tokens appearing in trajectory text were replaced with placeholders.

## License

CC BY 4.0 (`LICENSE`). Task prompts come from BFCL (Apache 2.0).

## Citation

```bibtex
@inproceedings{liu2026openclawbench,
  title     = {{OpenClawBench}: Benchmarking Process-side Anomalies in Real-world Agent Execution Trajectories},
  author    = {Liu, Yibing and Liu, Yangze and Yin, Xiaolong and Wang, Bin and Zhang, Chong and Yin, Hao and Han, Zhongyi},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track},
  year      = {2026}
}
```