Datasets:

Lightcap
/

agent-runtime-telemetry-small

pretty_name: Agent Runtime Telemetry Small
license: cc-by-4.0
language:
  - en
tags:
  - agent-runtime
  - agent-observability
  - llm-observability
  - mcp
  - tool-calling
  - runtime-telemetry
  - audit-trail
  - workflow-traces
  - parquet
size_categories:
  - 10K<n<100K
configs:
  - config_name: dataset_overview
    data_files:
      - split: train
        path: data/dataset_overview.parquet
  - config_name: operations
    data_files:
      - split: train
        path: data/operations.parquet
  - config_name: operation_events
    data_files:
      - split: train
        path: data/operation_events.parquet
  - config_name: artifact_records
    data_files:
      - split: train
        path: data/artifact_records.parquet
  - config_name: audit_records
    data_files:
      - split: train
        path: data/audit_records.parquet
  - config_name: tool_summary
    data_files:
      - split: train
        path: data/tool_summary.parquet
  - config_name: artifact_summary
    data_files:
      - split: train
        path: data/artifact_summary.parquet
  - config_name: daily_activity
    data_files:
      - split: train
        path: data/daily_activity.parquet

Agent Runtime Telemetry Small

Curated by Faruk Alpay.

Agent Runtime Telemetry Small is a compact tabular export of MCP-style agent execution telemetry. It is designed for dataset viewer inspection, lightweight agent observability experiments, tool-call reliability analysis, workflow trace summaries, and audit-trail research.

The dataset is intentionally small and row-oriented. Each table is stored as Parquet so the Hugging Face Dataset Viewer can display clean columns without requiring a SQLite client.

What It Contains

Config	Rows	Columns	Purpose
`dataset_overview`	7	6	Table inventory and export policy
`operations`	2,262	33	Tool execution records, status, stages, durations, and summarized result metadata
`operation_events`	9,903	13	Lifecycle events for operations
`artifact_records`	1,269	19	Forecast, state-decode, and training artifact index records
`audit_records`	14,053	17	Tool request/result audit rows with compact metadata
`tool_summary`	32	8	Aggregated tool reliability and latency statistics
`artifact_summary`	9	7	Aggregated artifact status and payload-size statistics
`daily_activity`	8	5	UTC daily activity counts across runtime tables

Privacy Boundary

This export does not upload the original SQLite databases and does not include raw nested payload_json bodies. Large JSON fields are represented with inspectable columns such as key lists, byte lengths, selected scalar status fields, and SHA-256 digests. Absolute local paths are reduced to path scope and file name columns.

Suggested Uses

compare agent tool success/error rates across runtime traces
inspect workflow latency and stage transitions
prototype LLM agent observability dashboards
analyze audit request/result volume without parsing full JSON logs
benchmark small-data telemetry pipelines that expect clean tabular inputs

Loading Example

from datasets import load_dataset

ops = load_dataset("Lightcap/agent-runtime-telemetry-small", "operations")
print(ops["train"][0])

summary = load_dataset("Lightcap/agent-runtime-telemetry-small", "tool_summary")
print(summary["train"].to_pandas().sort_values("operation_count", ascending=False).head())

Source

The rows were exported from local runtime SQLite stores into sanitized Parquet tables:

operation_state.sqlite3
artifact_store.sqlite3
audit_store.sqlite3

The export focuses on the operational shape of agent runtimes rather than application-specific content.