|
Download README.md from dynatrace-oss/SLS-Bench: direct link, hf CLI and curl.
- Browser
- Download file 2.76 kB
-
https://huggingface.co/datasets/dynatrace-oss/SLS-Bench/resolve/main/README.md
- Command line
-
hf download hf://datasets/dynatrace-oss/SLS-Bench/README.md
-
curl -L -o README.md https://huggingface.co/datasets/dynatrace-oss/SLS-Bench/resolve/main/README.md
2.76 kB
| license: cc-by-4.0 | |
| task_categories: | |
| - summarization | |
| - text-generation | |
| language: | |
| - en | |
| tags: | |
| - logs | |
| - observability | |
| - incident-response | |
| - benchmark | |
| - synthetic-data | |
| - llm-agents | |
| size_categories: | |
| - n<1K | |
| pretty_name: SLS-Bench | |
| configs: | |
| - config_name: default | |
| data_files: | |
| - split: test | |
| path: problems.jsonl | |
| # SLS-Bench: Synthetic Log Summarization Benchmark | |
| 90 incident log summarization problems for evaluating LLM agents. | |
| ## Files | |
| - `problems.jsonl` — 90 records with fields `id`, `question`, `answer`. The question is the same for every problem; only the underlying incident and reference answer differ. | |
| - `logs.zip` unzips to `logs/<id>.csv` — the log stream for each problem. Columns: `timestamp, level, message, trace_id, service, host`. | |
| ## Loading | |
| ```python | |
| from datasets import load_dataset | |
| import pandas as pd | |
| ds = load_dataset("dynatrace-oss/SLS-Bench", split="test") | |
| ex = ds[0] | |
| logs = pd.read_csv(f"logs/{ex['id']}.csv") # after unzipping logs.zip | |
| ``` | |
| or | |
| ```python | |
| import pandas as pd, mlcroissant | |
| u = lambda x: x.decode() if isinstance(x, bytes) else x | |
| ds = mlcroissant.Dataset( | |
| "https://huggingface.co/datasets/dynatrace-oss/SLS-Bench/resolve/main/croissant.json" | |
| ) | |
| problems = { | |
| u(r["problems/id"]): {k: u(r[f"problems/{k}"]) for k in ("question", "answer")} | |
| for r in ds.records(record_set="problems") | |
| } | |
| tasks = { | |
| pid: {**problems[pid], "logs": pd.read_csv(u(r["log_files/filepath"]))} | |
| for r in ds.records(record_set="log_files") | |
| if (pid := u(r["log_files/problem_id"])) in problems | |
| } | |
| print(tasks["01564611"]) # question, answer, logs | |
| ``` | |
| ## Statistics | |
| | | min | median | max | | |
| |---|---|---|---| | |
| | Log rows | 21,393 | 56,740 | 95,503 | | |
| | Log file size | 3.0 MB | 8.7 MB | 20.5 MB | | |
| | Reference answer length | 1,860 words | 2,418 words | 3,335 words | | |
| Total: 90 problems, ~830 MB of logs uncompressed. | |
| ## Responsible AI | |
| - **Limitations** — Evaluation-only; not for training, real-time triage, multi-modal evaluation, or operational certification. | |
| - **Biases** — Selection bias toward organizations that publish postmortems publicly; LLM-generated content inherits the generator's biases toward widely-documented technology stacks. | |
| - **Personal/sensitive information** — None. All identifier-shaped values in logs are synthetic. | |
| - **Use cases** — Validated for incident-summary quality scoring against a 23-LLM benchmark with calibrated LLM judges. | |
| - **Social impact** — Could reduce on-call burden if reliable; primary risks are over-reliance and treating synthetic-benchmark scores as operational readiness. | |
| - **Synthetic data** — Yes; logs and reference answers are fully synthetic, derived from public incident postmortems via an LLM pipeline. | |