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

```python
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.