adiven's picture
docs: rename dataset card to agent-reliability-bench
eabc80e verified
|
Raw History Blame Contribute Delete
5.54 kB
---
pretty_name: Agent Reliability Eval
configs:
- config_name: default
data_files:
- split: test
path: data/test.jsonl
tags:
- benchmark
- agent
- reliability
- gaia
- question-answering
---
# Agent Reliability Eval
本评测集在 **GAIA 的基础上筛选任务,并增添部分针对性任务,用于验证 Agent 的可靠性**。通过同题独立重复、已审核问法变体、工具扰动以及重复行为与中断恢复场景,研究 Agent 的正确性、一致性和鲁棒性。
Agent Reliability Eval is a reliability evaluation dataset built by selecting tasks from GAIA and adding targeted tasks for agent reliability validation. It supports repeated independent trials, reviewed question variants, tool perturbations, and studies of repetitive behavior and interruption/recovery.
## Composition / 组成
- **37 tasks** in the evaluation (`test`) split.
- **3 locally added tasks** provide targeted repetition and interruption/recovery scenarios; the remaining task records form the selected subset.
- **185 reviewed question variants**: five variants for each task.
- **6 attachments**, with relative paths and SHA-256 hashes.
- Explicit reference answers, scored by the associated benchmark's deterministic GAIA-style QA scorer.
Locally added task IDs: `attachment-repeat-interruption-001`, `path-review-interruption-001`, `tool-repeat-interruption-001`. Their levels are local assignments, not official GAIA difficulty annotations.
## Purpose and construction / 用途与构建
GAIA-derived tasks were filtered and curated through the Insight reliability evaluation workflow. Targeted tasks were added to exercise repeated tool calls, repetitive text/path review, and interruption/recovery responses. This is a curated experimental set, not a random representative sample of GAIA; selection involved observations from earlier local experiments, which can introduce selection bias.
Question variants preserve the intended task and reference answer. Repeats and tool perturbations are configured by the runner rather than duplicating task rows. Task construction and actual RAS detection are separate observations: detectors can trigger on tasks beyond the three intentionally constructed cases.
Reliability dimensions and aggregate definitions depend on the benchmark protocol. This release does not claim measured performance or a universal aggregate reliability score.
## Code / 代码
[agent-reliability-bench on GitCode](https://gitcode.com/weixin_46351149/agent-reliability-bench)
The code remains on GitCode. This repository distributes dataset assets independently of the Insight authoring installation. Users do not need Insight installed to consume this snapshot.
## Files / 文件
| File | Purpose |
|---|---|
| `data/test.jsonl` | One task per line for HF Datasets loading and authorized viewing |
| `dataset.json` | Native snapshot with questions, reference answers and dataset fingerprint |
| `variants.json` | Reviewed variants, explicit styles and matching dataset fingerprint |
| `attachments/` | Task attachments, preserving referenced relative paths |
| `release.json` | Release identity, counts and per-file SHA-256 inventory |
`data/test.jsonl` and `dataset.json` contain identical task records. They are two representations of one task set.
## Loading / 加载
After receiving access and authenticating, use HF Datasets:
```python
from datasets import load_dataset
tasks = load_dataset(
"adiven/agent-reliability-bench",
split="test",
revision="REPLACE_WITH_FULL_RELEASE_COMMIT_SHA",
)
```
For all assets, including question variants and attachments:
```python
from huggingface_hub import snapshot_download
root = snapshot_download(
repo_id="adiven/agent-reliability-bench",
repo_type="dataset",
revision="REPLACE_WITH_FULL_RELEASE_COMMIT_SHA",
)
```
Replace the revision placeholder with the uploaded release commit. Resolve attachment paths relative to `root`. Verify `release.json` and preserve snapshot contents. Reference answers are scorer inputs; Agent input should contain only the question and task attachments. Do not modify the shared HF cache in place.
Automatic HF prepare/download integration in the benchmark is planned. The current implementation accepts an authorized local resource snapshot; see the GitCode repository's `docs/data-setup.md`. Pin the code version and data commit for reproducible comparisons.
## Access, provenance and use / 访问与来源
Upstream: [gaia-benchmark/GAIA](https://huggingface.co/datasets/gaia-benchmark/GAIA).
This release contains GAIA-derived material. GAIA's stated access conditions require resharing on the HF Hub to remain in a gated or private repository and discourage crawlable redistribution. This repository is initially private; a later access-request release must preserve applicable upstream conditions. Source tasks and attachments retain their upstream terms; this release does not grant a blanket replacement license for third-party content.
A source-snapshot fingerprint is retained for auditability. Internal experiment identifiers and historical operational records are omitted. This sanitization changes dataset/variant package fingerprints; task text, reference answers, attachment bytes and task IDs are preserved. Fingerprints identify this standalone release and can differ from an earlier local Insight import.
No evaluation trajectories, model credentials, private runtime configuration or previous experiment outputs are included.