README.md

---
license: cc-by-4.0
language:
- zh
- en
size_categories:
- 1K<n<10K
task_categories:
- other
tags:
- agent-evaluation
- llm-as-judge
- conformal-prediction
- benchmark
- mobile-agent
- browser-agent
- desktop-agent
pretty_name: GeneralAgentBench
configs:
- config_name: default
  data_files:
  - split: test
    path: general_agent_bench.jsonl
---

# GeneralAgentBench

GeneralAgentBench is a 1,400+ task benchmark for evaluating whether general-purpose AI agents have **genuinely completed** a task, spanning **Mobile / Browser / Desktop** environments. It is the evaluation resource accompanying an anonymous NeurIPS 2026 Evaluations & Datasets Track submission (Submission 173, *AgentJudge*). This release is fully anonymized for double-blind review.

Each task provides a natural-language instruction plus a list of verification
checkpoints. This release is a **task catalog**: it does not include agent
trajectories, independent human verdicts, checkpoint-level labels, or the
experimental calibration/test assignment used for the paper's reported
evaluator metrics.

## Dataset at a glance

| Property | Value |
|---|---|
| Total tasks | 1,448 |
| Environments | `desktop_linux` (1,023), `mobile_android` (403), unspecified (22) |
| Difficulty | easy (739), medium (512), hard (197) |
| Priority | P0 (487), P1 (693), P2 (197), P3 (71) |
| Avg. checkpoints / task | 4.23 (every task has ≥1) |
| Tracks | `skill` (963), `mobile` (306), `mobile_agent` (97), `general` (82) |
| Language | Chinese instructions with some English; task targets reference public websites |
| License | CC BY 4.0 |

> Note on anonymization: all references to specific commercial products, services, and internal codenames have been replaced with neutral placeholders (e.g. `paymentapp`, `mapapp`, `cloudplatform`, `doccollab`, `teamapp`, `shoppingapp`, `skillkit`). Task logic and difficulty are preserved. Publicly known third-party websites used as task targets are retained, as they do not identify the authors.

## Data fields

Each line in `general_agent_bench.jsonl` is one task:

| Field | Type | Description |
|---|---|---|
| `uid` | string | Globally unique id (`<track>/<id>`) |
| `track` | string | Source track: `skill` / `mobile` / `mobile_agent` / `general` |
| `category` | string | Top-level category (anonymized) |
| `subcategory` | string | Full category path (anonymized) |
| `id` | string | Original task id within its track |
| `question` | string | The task instruction given to the agent |
| `reference_answer` | string | Short reference / expected-outcome note (may be empty) |
| `checkpoints` | list[string] | Human-defined verification conditions for success |
| `priority` | string | P0–P3 |
| `difficulty` | string | easy / medium / hard |
| `environment` | string | `desktop_linux` / `mobile_android` / empty |
| `execution_mode` | string | e.g. `one_shot`, multi-turn variants |
| `timeout` | int | Suggested wall-clock budget (seconds) |
| `step_budget` | int | Suggested max agent steps |
| `token_budget` | int | Suggested token budget |
| `precondition_files` | string (JSON) | Pre-staged input files under `assets/` (JSON-encoded list) |
| `tags` | list[string] | Optional tags |
| `extra_json` | string (JSON) | Any additional track-specific fields, JSON-encoded |

Pre-staged input files referenced by `precondition_files` are provided in the `assets/` directory (synthetic CSVs, sample documents, generic test images — no personal data).

## Usage

```python
from datasets import load_dataset

ds = load_dataset("agentjudge-anon/GeneralAgentBench", split="test")
print(ds[0]["question"])
print(ds[0]["checkpoints"])

# nested fields are JSON-encoded strings:
import json
pre = json.loads(ds[0]["precondition_files"])
extra = json.loads(ds[0]["extra_json"])
```

## Intended use and evidence scope

The released task catalog can be used to construct agent or evaluator studies,
inspect task/checkpoint coverage, and test data-loading code. Reproducing the
paper's human-agreement, error-rate, conformal-coverage, or routing results
additionally requires trajectories, independent reference labels, label
nonconformity scores, and the original experimental split assignment; those
artifacts are **not included in this release**. Task availability must not be
mistaken for evaluator validation.

Hugging Face exposes the complete JSONL as one `test` configuration. The file
does not contain a `split` field and therefore is not a
development/calibration/test split manifest.

## Repository-evidenced composition

The 1,448 released rows are aggregated from four tracks and normalized to the
schema documented above:

| Track | # tasks | Origin | Environment |
|---|---|---|---|
| `skill` | 963 | Skill/capability regression suites | `desktop_linux` |
| `mobile` | 306 | Mobile UI task suite | `mobile_android` |
| `mobile_agent` | 97 | Mobile agent task suite | `mobile_android` |
| `general` | 82 | General cross-environment tasks | mixed |

The file directly supports exact row, track, environment, difficulty, priority,
and checkpoint counts; `uid` uniqueness; field-type checks; and file hashes.

The task-level `priority` field contains `P0`–`P3`; it is not a per-checkpoint
MUST/NICE/OPT label. The `difficulty` field contains easy/medium/hard labels,
but the component values `V`, `K`, and `S` are not distributed, so the
log-linear difficulty formula described in the paper cannot be recomputed from
this release alone.

## Construction and annotation records

The accompanying paper describes a three-stage construction protocol. This
public task-only release does not contain per-annotator votes, adjudication
records, reference-agent dry-run traces, or contamination-comparison outputs.
Consequently, annotator counts, agreement statistics, human-vs-LLM
responsibility, and contamination rates are historical manuscript claims
rather than measurements independently auditable from the released JSONL.

## Release validation

`RELEASE_MANIFEST.json` records machine-checkable counts, field names, and
limitations. `checksums.sha256` contains SHA-256 hashes for the distributed
task file, Croissant metadata, and manifest. These files validate the release
contents; they do not reconstruct the paper's experimental splits or labels.

## Limitations

- Instructions are predominantly in Chinese.
- Agent trajectories, screenshots, independent human labels, calibration
  scores, and original experimental split assignments are not released here.
- Anonymized product placeholders mean some instructions read generically (e.g. "the payment app") rather than naming a specific service.
- Task-level priority is not checkpoint-level priority annotation.
- The release does not by itself reproduce the paper's evaluator metrics.

## Citation

Anonymous. *AgentJudge: Conformal Evaluation of General-Purpose AI Agents via Multi-Modal Evidence Fusion, Six-Dimensional Scoring, and Dual-Track Adaptive Routing.* NeurIPS 2026 Evaluations & Datasets Track, Submission 173 (under review). Citation details will be completed upon acceptance.