Skip to content

Published Harbor datasets

Results cover the six native pipelines, Tasksmith and all 14 research recipes. Historical evidence and the newer publication checks are reported separately below.

Browse the HuggingEnvs collection. The native datasets retain their existing owners.

Native pipelines

600 task entries across 6 earlier datasets. Recovered from cached Hub manifests and local publication stagings on 2026-09-15. These are historical snapshots, not a fresh Hub recount. Older revisions and duplicate stagings are excluded; cross-pipeline content is not deduplicated.

Pipeline Tasks Recovered validation evidence Dataset and evidence
pr_diff 181 181 listed; no per-task validation in this snapshot. Earlier 100-task release reported oracle passes. Dataset · Manifest
pr_runtime 100 100 oracle/tracked passes; 88 clean commands; 87 with a regression guard. Dataset · Manifest
commit_runtime 100 100 generation-time verified stamps. Full 100-task oracle gate not recovered; older 52-task gate is separate. Dataset · Manifest
code_instruct 100 100 emitted after generation checks; no separate 100-task Harbor gate recovered. Dataset · Local evidence
equivalence_tests 100 100 emitted after generation checks; no separate 100-task Harbor gate recovered. Dataset · Local evidence
cve_patches 19 19 listed; validation fields absent. A 19-task oracle or independent quality gate was not recovered. Dataset · Manifest

These tasks have historical evidence scopes, not retrospectively assigned verified labels. In particular, the earlier 52-task commit-runtime gate does not validate the later 100-task dataset. Read the native results and solver samples.

Tasksmith and research recipes

1,330 tasks across 15 datasets. Each dataset contains complete Harbor task directories, archives and a registry pinned to its artifact revision.

Pipeline Tasks Evaluation labels Dataset and pinned manifest
swe-smith 100 100 unverified Dataset · Manifest
r2e 100 100 unverified Dataset · Manifest
swe-gen 100 100 unverified Dataset · Manifest
swe-next 100 100 unverified Dataset · Manifest
r2e-gym 100 100 unverified Dataset · Manifest
scaler 100 100 unverified Dataset · Manifest
endless-terminals 100 100 unverified Dataset · Manifest
cli-gym 25 25 unverified Dataset · Manifest
swe-flow 100 2 needs repair; 98 unverified Dataset · Manifest
seta-seed2synth 100 100 unverified Dataset · Manifest
seta-evol 100 100 unverified Dataset · Manifest
tmax 55 55 unverified Dataset · Manifest
terminalworld 100 3 needs repair; 97 unverified Dataset · Manifest
dataarc 100 100 unverified Dataset · Manifest
tasksmith 50 50 verified Dataset · Manifest

What the labels establish

The Tasksmith and research-recipe release contains 50 verified, 5 needing repair and 1,275 unverified tasks. These totals exclude the historical native inventories above. Tasksmith's verified cohort came from an assisted campaign; this does not claim unattended conversion. Two SWE-flow instruction issues and three TerminalWorld verifier gaps remain explicitly diagnosed. Each dataset manifest supplies task-level labels, diagnostics and evidence scope.

Publication checks for those 15 datasets compared 214,097 file identities and parsed every selected task with Harbor. This establishes artifact integrity and format, not semantic quality of every task. See evaluation labels, yield and cost, and how to publish.