---
license: mit
task_categories:
  - text-classification
language:
  - en
tags:
  - agents
  - multi-agent-systems
  - failure-analysis
  - mast
  - evaluation
  - arxiv:2503.13657
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files: "*.jsonl"
---

# MAST Failure-Mode Gold Set

25 **human-verified** multi-agent failure traces, each labelled with one of the 14 MAST
failure modes (Cemri et al., "Why Do Multi-Agent LLM Systems Fail?",
[arXiv:2503.13657](https://arxiv.org/abs/2503.13657)). Built as the ground-truth set for
evaluating the MAST classifier in
[adk-agent-playground](https://github.com/barissozudogru/adk-agent-playground) — i.e. for
Cohen's κ against an LLM judge, not for training.

## What's in a row

| field | meaning |
|---|---|
| `id` | stable id, e.g. `gold-001` |
| `pipeline_name` / `agent_name` | which agent produced the output |
| `user_query` | the request the agent was given |
| `output_snippet` | the agent output exhibiting (or not) a failure |
| `gold_label` | the human-assigned MAST mode, e.g. `FM-1.1` |
| `gold_confidence` | annotator confidence in [0, 1] |
| `rationale` | one sentence on why the label applies |

25 rows across 14 modes (roughly 2 per mode — built for balanced κ evaluation, not training).
The three MAST categories: FC1 system design (FM-1.x), FC2 inter-agent misalignment (FM-2.x),
FC3 task verification (FM-3.x).

## Honest scope — this is an EVAL set, and it is small

25 examples over 14 classes is **far too small to train** a classifier. A documented reference
point: a char-n-gram nearest-centroid baseline gets **9/25 = 36% leave-one-out accuracy** here
(Wilson 95% CI ≈ [20%, 55%]), against an 8% majority-class baseline and ~7% chance. The interval
sits clear of chance but is wide — which is the honest reading on n = 25, and exactly why a bare
point estimate would overstate it. Real signal, nowhere near deployable: treat this as a hardness
probe and a κ ground-truth set, not a training corpus. The traces are
hand-authored to exemplify each mode (some paraphrase real failures); they are illustrative, not
a random sample of production failures, so do not read a class prevalence off them.

## Intended use

- Ground truth for measuring a failure-mode classifier's agreement (Cohen's κ, per-mode
  precision/recall) against human labels.
- A worked illustration of each of the 14 MAST modes for documentation and onboarding.

## Related artifacts

- Space — [harness-science-evolution](https://huggingface.co/spaces/barissozudogru/harness-science-evolution): the live ablation + governed-evolution demo from the same project.
- Sibling datasets — [belnap-contested-questions](https://huggingface.co/datasets/barissozudogru/belnap-contested-questions) · [swe-bench-mini](https://huggingface.co/datasets/barissozudogru/swe-bench-mini).
- Source — [github.com/barissozudogru/adk-agent-playground](https://github.com/barissozudogru/adk-agent-playground).

## License

MIT, matching the source project.