MonikaDvorackova's picture
Add synthetic agent reliability trace dataset
096a26b unverified
|
Raw
History Blame Contribute Delete
3.11 kB
metadata
license: apache-2.0
language:
  - en
pretty_name: Agent Reliability Traces
size_categories:
  - n<1K
task_categories:
  - text-classification
tags:
  - agents
  - agent-evaluation
  - reliability
  - ai-safety
  - responsible-ai
  - synthetic
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train.jsonl

Agent Reliability Traces

A small synthetic dataset of observable AI agent execution traces annotated with reliability and failure-mode signals.

The dataset accompanies the Agent Reliability Lab Hugging Face Space.

Dataset purpose

The dataset is designed for:

  • prototyping agent-trace evaluation
  • testing deterministic reliability heuristics
  • experimenting with failure-mode classification
  • evaluating tool-use trajectories
  • educational and portfolio use

It is not intended as a production benchmark.

Structure

Each record contains:

Field Description
id Unique trace identifier
goal Task assigned to the agent
trace Observable execution trace
risk_label Coarse reliability-risk label
failure_mode Primary detected failure mode
has_verification Whether explicit verification is present
has_tool_failure Whether tool execution failures are present
has_loop Whether repeated tool behaviour is present
has_final_answer Whether an explicit final answer is present
contains_overconfidence Whether overconfident language is present
notes Annotation explanation

Risk labels

Current labels include:

  • low
  • moderate
  • elevated
  • high

Failure modes

Examples include:

  • none
  • tool_failure
  • loop_and_tool_failure
  • missing_verification
  • unsupported_answer
  • missing_final_answer
  • overconfidence

Data creation

All examples are synthetic and manually constructed for this project.

They are not production traces and do not contain hidden chain-of-thought.

The dataset includes only observable execution-style events such as plans, tool calls, observations, errors and final answers.

Intended use

This dataset can be used to explore:

  • agent reliability analysis
  • failure-mode detection
  • trace classification
  • tool-use evaluation
  • verification behaviour
  • loop detection
  • structured agent monitoring

Relationship to Agent Reliability Lab

The associated Agent Reliability Lab uses deterministic heuristics to analyze observable execution traces.

This dataset provides controlled examples that can be used to test and extend that approach.

Limitations

The dataset is deliberately small and synthetic.

The annotations should not be treated as authoritative measures of agent safety or correctness.

The current traces use a simplified human-readable format and do not represent the event schemas of specific agent frameworks.

Future work

Potential extensions include:

  • larger trace collections
  • structured JSON trajectories
  • LangGraph traces
  • tool-call graphs
  • multi-agent interactions
  • human annotations
  • learned failure-mode classifiers
  • independently collected evaluation traces

License

Apache-2.0