sandboxapis/olympus-labs-evals

olympus-labs answer-key evals

Repository id: sandboxapis/olympus-labs-evals.

Dataset Summary

53 questions an agent can answer by calling real API shapes: GitHub, GitLab, Jira, Linear and Slack. Each row carries the question, the request that answers it, and the value the answer must be. The requests go to SandboxAPIs pinned snapshot hosts, which are read-only replicas of those APIs preloaded with one simulated engineering company, olympus-labs. A pinned host returns the same bytes on every request, so a row's expected answer is as true next year as it is today.

All 53 rows follow one incident across the five services: the bug report in the tracker, the pull request that fixed it, its review, the branch it merged from, and the chat thread where it was called resolved. The identifiers agree from one service to the next, so some rows can only be answered by reading two responses and joining them.

This is version olympus-labs-g12-2026-09: generation g12 of the data set, compiled from template hello-20, artifact sha256 c27f762cc1d2656fca27e2a3ffc8c7742e41bca227cc7d36efc4cf0a8bf31d39.

from datasets import load_dataset
import json

rows = load_dataset("sandboxapis/olympus-labs-evals", split="train")
row = rows[0]
expected = json.loads(row["expected"])  # {"path": "number", "match": "equals", "value": 53}

Supported Tasks

• Tool-using agent evaluation. Give the agent question and an HTTP tool (or the SandboxAPIs MCP server), let it choose its own calls, and score its final answer against expected. The tool_calls column is one correct route to the answer, not the only one.
• Answer extraction from API responses. Fetch tool_calls yourself, hand the responses to the model with question, and score what it reads out of them. This isolates reading from planning.

Scoring is exact and needs no judge model. expected.match is one of:

match	the answer is correct when
equals	it equals expected.value (a string, an integer or a boolean)
count	it is a list whose length is expected.value
contains	it is a string that contains expected.value

expected.path is where the value sits in the response to request_url, as a dotted path in which numbers index arrays and "" is the whole document.

difficulty grades the row: read is one field off one response, join means two responses have to agree and the second call is only reachable once the first has been read, and reason means the answer is computed from what came back rather than copied out of it.

Two zero-dependency harnesses that score either the data set or your model, run.mjs and run.py, are in driftwork-llc/sandboxapis-examples.

Dataset Structure

Data Fields

field	type	what it holds
id	string	Stable row id, for example gh-pull-number. Never reused for a different question.
provider	string	github, gitlab, jira, linear or slack.
question	string	The question, in plain English, as an agent would be asked it.
expected	string	JSON: {"path", "match", "value"}, described above. A string because value changes type between rows.
request_url	string	The request whose response holds the answer.
request_method	string	GET, or POST for the Linear GraphQL rows.
tool_calls	string	JSON: every request the answer needs, in order, as {"method", "url", "body"}. body is the GraphQL document on Linear rows and null elsewhere.
difficulty	string	read, join or reason.
guarded_by	string	The test in the SandboxAPIs engine repository that fails if this answer changes. That repository is private; the citation shows that each answer is held by a test rather than typed from memory.
pin_host	string	The pinned snapshot host that request_url points at.

Example row

{
  "id": "gh-pull-number",
  "provider": "github",
  "question": "Which pull request number in olympus-labs/parthenon carried the fix for the dropped-events incident?",
  "expected": "{\"path\":\"number\",\"match\":\"equals\",\"value\":53}",
  "request_url": "https://gh-2026-03-g12.snap.sandboxapis.dev/repos/olympus-labs/parthenon/pulls/53",
  "request_method": "GET",
  "tool_calls": "[{\"method\":\"GET\",\"url\":\"https://gh-2026-03-g12.snap.sandboxapis.dev/repos/olympus-labs/parthenon/pulls/53\",\"body\":null}]",
  "difficulty": "read",
  "guarded_by": "packages/conformance/src/parity/pinned-cross-category.test.ts#closed_by_pull resolves to the closing PR on the GITHUB host, via the tracker's branchName",
  "pin_host": "gh-2026-03-g12.snap.sandboxapis.dev"
}

Data Splits

One split, train, with all 53 rows. There is no held-out split: every answer can be read from a public host, so a held-out split would not stay held out. Treat the set as an evaluation set and do not train on it.

provider	rows	pinned host
github	12	gh-2026-03-g12.snap.sandboxapis.dev
jira	11	jira-v3-g12.snap.sandboxapis.dev
slack	14	slack-2026-08-g12.snap.sandboxapis.dev
gitlab	6	gl-v4-g12.snap.sandboxapis.dev
linear	10	linear-2026-08-g12.snap.sandboxapis.dev

Dataset Creation

Source data

The data is generated, not recorded. A deterministic simulation runs a fictional company's engineers over 90 days of simulated time, and a compiler renders the result in each provider's API shape. Nothing in it came from a real account, a real repository or a real person.

There is one fictional company, olympus-labs, and its people and projects carry Greek-mythology names (athena, tiresias, parthenon) so that nobody mistakes a row for a real organisation's data.

Annotations

Each expected value was read off the pinned host and is held there by a test in the engine repository, named in guarded_by. A row nobody can cite does not go in the set.

Considerations for Using the Data

• Read-only hosts. The pinned hosts accept reads only. A write returns the provider's own error status and body with a plain explanation, so an agent under test cannot change the answers for anyone else.
• Rate limit. Anonymous callers get 60 requests an hour per IP. Scoring every row with the reference tool_calls takes 12 requests, because rows share responses. A free key from sandboxapis.dev/login raises the limit to 600 an hour; send it as Authorization: Bearer <key>.
• Every answer is checkable live. Fetch request_url and read expected.path out of it. If the value differs, the row or the host is wrong, and we want to hear about it at sandboxapis.dev/feedback.
• Pinned, not live. The -g12 hosts are frozen snapshots. SandboxAPIs also serves live hosts that roll forward daily; do not point this set at them, because their answers move.
• Contamination. The rows and their answers are public. A model may have seen them, so report a score on this set alongside your own held-back tasks rather than instead of them.

Licensing Information

MIT. The hosted service and the engine that generates the data set are separate and are not covered by this licence.

Citation Information

@misc{sandboxapis_olympus_labs_evals_2026,
  author       = {{SandboxAPIs team, Driftwork LLC}},
  title        = {olympus-labs answer-key evals},
  year         = {2026},
  version      = {olympus-labs-g12-2026-09},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/sandboxapis/olympus-labs-evals}}
}