File size: 3,052 Bytes
5495b89 | 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 | ---
license: agpl-3.0
pretty_name: ToolGate-Bench
language:
- en
task_categories:
- text-classification
size_categories:
- n<1K
tags:
- agents
- tool-use
- agent-safety
- guardrails
- benchmark
- snapkitty
configs:
- config_name: default
data_files:
- split: v1
path: data/v1.jsonl
- split: heldout
path: data/heldout.jsonl
---
# ToolGate-Bench
A small, deterministic benchmark for **tool-call gating**: given raw LLM output, should an agent harness `ACCEPT` it, require human `APPROVAL_REQUIRED`, or `REJECT` it? It targets the failure that matters most for agents, a risky action slipping through a gate (an *unsafe pass*).
Reference gate: [Snapkitty/snapkitty-nemotron-harness](https://huggingface.co/Snapkitty/snapkitty-nemotron-harness) (TypeScript, pure function, ~24 µs per decision).
## Splits
| Split | Cases | Purpose |
|---|---:|---|
| `v1` | 78 | Canonical, combined, obfuscated and undeclared tool tokens, denied shell commands, benign outputs |
| `heldout` | 18 | Frozen generalization set: new spellings, commands outside the denylist, dangerous strings inside harmless code |
Each row: `id`, `category`, `output` (the model text), `expected` (`ACCEPTED` / `APPROVAL_REQUIRED` / `REJECTED`). Expected labels follow the [trust deed](https://huggingface.co/Snapkitty/snapkitty-nemotron-harness/blob/main/trust-deed.json) of the reference harness.
```python
from datasets import load_dataset
ds = load_dataset("Snapkitty/toolgate-bench")
print(ds["v1"][0])
```
## Metrics
- **Accuracy**: exact match with `expected`.
- **Unsafe pass rate**: share of non-`ACCEPTED` cases the gate accepted. Lower is better; this is the headline safety number.
- **False reject rate**: share of benign cases the gate blocked.
## Reference results
| Gate | Split | Accuracy | Unsafe passes | False rejects |
|---|---|---:|---:|---:|
| snapkitty-nemotron-harness v0.2.0 | v1 | 39.7% | 82.5% | 0% |
| snapkitty-nemotron-harness v0.3.0 | v1 | **100%** | **0%** | **0%** |
| snapkitty-nemotron-harness v0.3.0 | heldout | 66.7% | 28.6% | 50% |
The `v1` split was written together with the v0.3.0 rules, so `heldout` is the fairer measure of a new gate. Submit your gate's numbers on both splits.
## Citation
```bibtex
@misc{snapkitty_toolgate_bench_2026,
title = {ToolGate-Bench: Deterministic Evaluation of Tool-Call Gating for LLM Agents},
author = {{Snapkitty Collective LLC}},
year = {2026},
howpublished = {\url{https://huggingface.co/datasets/Snapkitty/toolgate-bench}}
}
```
## License
Licensed under **AGPL-3.0**. Full text: [LICENSE](LICENSE).
### 💼 Commercial License
SnapKitty code is free and open under **AGPL-3.0** for open-source use. Building a commercial product or service? A **proprietary commercial license** from Snapkitty Collective LLC lets you ship SnapKitty code without the AGPL's source-sharing and network-use obligations.
**[→ Get a commercial license](mailto:A.parr@belespritdaccord.uk?subject=Commercial%20license:%20toolgate-bench)** · A.parr@belespritdaccord.uk
|