Datasets:

TraceLite-AI
/

TraceTrial-Bench

language:
  - en
license: mit
size_categories:
  - n<1K
pretty_name: TraceTrial-Bench
task_categories:
  - other
tags:
  - benchmark
  - agents
  - harness
  - reliability
  - failure-attribution
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/tasks.jsonl
  - config_name: traces
    data_files:
      - split: test
        path: traces/index.jsonl

TraceTrial-Bench

Trial by Trace: Benchmarking AI Agent Harnesses for Reliability and Failure Attribution

Paper · Interactive report / Space · Code · Download all tasks

Dataset contents

The final 95 executable task packages cover 40 mechanism nodes across 7 reliability dimensions. Each package contains its prompt, workspace inputs, runtime/simulator configuration, and Harness/Agent scoring files. The dataset also includes 3,420 model execution trajectory packages covering 95 tasks × 9 models × 4 harnesses; see Execution trajectories.

Dimension	Tasks
Context	15
Execution	27
Governance	9
Lifecycle	19
Observability	7
Tools	17
Verification	1
Total	95

Task/final.manifest is the authoritative task order, matching the heatmap's left-hand labels from top to bottom on page 45, Figure 13 of the supplied paper PDF.
Task/ contains the complete task directories. Task IDs and task input README files are preserved.
data/tasks.jsonl provides one metadata/prompt row per task for the Dataset Viewer; paper_order is one-based. This index is not a replacement for the complete task packages.
TraceTrial-Bench-tasks.zip contains Task/ and LICENSE, ready to extract into the code repository.
checksums.sha256 records the SHA-256 of every task file. release.json records the source revision and archive digest.

Execution trajectories

The traces/ release contains 3,420 runs (95 tasks × 9 models × 4 harnesses), distributed as 95 per-task .tar.gz archives. Each archive contains all 36 model/harness run directories for that task, retaining execution traces, available workspace evidence, and scoring artifacts.

traces/index.jsonl: one row per run, with task/model/harness identity, execution completeness, historical result status, final combined score, archive location, and original score/result hashes.
traces/task_name_mapping.tsv: original-to-current task directory names. Embedded historical IDs and paths remain unchanged.
traces/release.json: provenance, counts, archive hashes, and publication redaction inventory.
traces/SHA256SUMS: checksums for the trace release files.
traces/README.md: extraction instructions and interpretation of execution versus scoring status.

Preview the run index:

from datasets import load_dataset

runs = load_dataset("TraceLite-AI/TraceTrial-Bench", "traces", split="test")
print(runs[0]["final_score"])

Download all trace archives (or select a task archive):

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="TraceLite-AI/TraceTrial-Bench",
    repo_type="dataset",
    local_dir="TraceTrial-Bench-data",
    allow_patterns=["traces/**"],
)

Historical result.json may still say run_complete or contain a pending evaluation even when a final artifacts/llm_judge/combined_score.json exists. The index explicitly separates these fields and uses the latter for final_score. Some runs have incomplete execution traces; these are identified by trace_complete, not silently excluded. Matched provider configuration values and recognizable access-token strings are redacted in the publication copy; source trajectories are retained locally. This upload does not rerun agents or judges.

Use

Preview metadata and prompts:

from datasets import load_dataset

tasks = load_dataset("TraceLite-AI/TraceTrial-Bench", split="test")
print(tasks[0]["task_id"])

Download the complete task packages:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="TraceLite-AI/TraceTrial-Bench",
    repo_type="dataset",
    local_dir="TraceTrial-Bench-data",
    allow_patterns=["Task/**", "LICENSE", "release.json", "checksums.sha256"],
)

The executable runner and installation instructions are in the code repository. From that repository, with Task/ in place and a configured model registry, select the full benchmark using --task-manifest Task/final.manifest.

Provenance and scope

The task-package portion of this dataset is exported from GitHub commit 6bd9524c7e89d4e6cc612d6f720abd24ccb44229. File bytes come from that committed Task/ tree, excluding untracked files, Python caches, credentials and run outputs. Task contents and scoring criteria are unchanged by this export. The test split contains the benchmark's full published selection; no train/validation split is introduced.

The task publication verifies packaging and order. The trajectory publication packages existing local run evidence and saved final scores, with its own provenance and redaction inventory in traces/release.json. Neither publication constitutes a new model evaluation or a rerun of the paper's experiments. The task environments are constructed benchmark scenarios; some task packages include source-document excerpts with their existing notices retained.

License

The repository license is MIT. Existing notices in bundled reference material are retained.

Citation

@misc{tracelite2026trialbytrace,
  title  = {Trial by Trace: Benchmarking AI Agent Harnesses
            for Reliability and Failure Attribution},
  author = {Cao, Zhiyong and Zhu, Hengjie and Wang, Yongxin and Li, Xuchen and
            Li, Jintao and Liu, Shuxia and Liu, Dunqiang and Wang, Bo and
            Zhang, Kairan and Lin, Cong and Yu, JiangDi and Liu, Siyuan and
            Yao, Sijia and Wei, Yi-Liang and Ma, Lingxu and Wang, Aijia and
            Chen, Anxin and Wei, Zixian and Ruan, Xiang and Xu, Ruizhi and
            Shi, Ruqian and Xiao, Baihua and Zhou, Sicheng},
  year   = {2026},
  eprint = {2610.32285},
  archivePrefix = {arXiv},
  primaryClass = {cs.AI},
  url    = {https://arxiv.org/abs/2610.32285}
}