Datasets:

keppy
/

evalroute-tasks

license: mit
language:
  - en
task_categories:
  - text-classification
pretty_name: evalroute-tasks
size_categories:
  - n<1K
tags:
  - evalroute
  - task-routing
  - coding-agents

evalroute-tasks

A small human-labeled corpus for training and evaluating evalroute's task→lane router. Each row is one real coding/research/writing task with the lane a human asserts it belongs to. Tier 1 (task → lane) is the only required label today; the schema leaves room for tier 2 (task → arm outcome) and tier 3 (task → done artifact) so contributors can backfill later without a format break.

Status: zero to one. One contributor, 54 rows, four of nine lanes. Per lane: routine-coding 30, dl-ml-research-engineering 13, alignment-reasoning 10, prose 1; hard-agentic-coding, long-doc-reading, web-research, math-first-principles and orchestration have none yet. 24 rows are real tasks pinned in one dispatcher's ledger (mostly coding briefs), 30 are hand-written tier-a taskset rows. n is small and the mix reflects one person's work — that is the point of publishing it: your work shapes your router, and the shared encoder is only as broad as this file. The trained encoder is keppy/evalroute-lane-encoder; the outcomes flywheel (lane/arm/verdict, no text) is keppy/evalroute-flywheel.

Tiers

Tier	Question	Keys	Required
1	which lane?	id, text, label, source, contributor, week	yes
2	which arm, and did it work?	arm, verdict, method, n	optional (all-or-none)
3	how was done judged?	checker / rubric / reference (at most one)	optional

Row example

{"id": "keppy:a3f9c1", "text": "add a CLI verb that reads the ledger and prints per-lane counts",
 "label": "routine-coding", "source": "ledger-pinned", "contributor": "keppy", "week": "2026-W40",
 "arm": "z-ai/glm-5.3-flash@medium", "verdict": "pass", "method": "measured", "n": 10,
 "checker": null, "rubric": null, "reference": null}

The publish gate (three steps, all must pass)

Mechanical validation: python scripts/validate.py tasks/*.jsonl checks id/text/lane/source/contributor/week format, tier-2 and tier-3 coherence, and a secret/PII/hostname regex gate over text and tier-3 strings.
Denylist: if private_terms.txt exists in the repo root (git-ignored) or you pass --private-terms <file>, any case-insensitive whole-word hit fails the row. Keep employer/project terms you must not publish in that file.
Human review: run python scripts/review.py tasks/<you>.jsonl, read every numbered row, compute the batch sha, add reviewed: <sha> <date> <contributor> to REVIEWED.md, then open a PR. scripts/upload.py refuses to touch Hugging Face unless REVIEWED.md contains a line for every file's sha.

How to contribute

Fork/branch, add your own file tasks/<your-handle>.jsonl (one JSON row per line; the file stem and row-id prefix must be your handle).
Run python scripts/validate.py tasks/<your-handle>.jsonl — it must exit 0.
Run python scripts/review.py tasks/<your-handle>.jsonl, actually read the rows, and add the printed reviewed: line to REVIEWED.md.
Open a PR. A reviewer re-runs validate and diff-checks your rows.

What is not here

No paraphrases: train F showed paraphrases that outnumber real rows pull the classifier toward the paraphrase dialect. No ledger exports without review — a row lands here only after a human decided the text is safe to publish. No personal or employer details: names, internal hostnames, credentials, and project-denialist terms are caught by the gate, but judgment is still yours.