affine-mswea

1,150 coding-agent trajectories on 1,129 real pull-request tasks, converted to the format of albedo's mini-coder source (mini-swe-agent: one ```bash command per assistant turn, <returncode>/<output> observations, mini-coder's system prompt and task template).

• Origin: Affine's public training corpus (https://data.affine.io, view duel_turns@v4, corpus epoch 84).
• Machine: Affine's mini-swe-agent text harness (each command runs in dash as /bin/sh -c, off a terminal). Albedo tells machines apart by source name, so each kind of machine is its own dataset.
• Repository root: rewritten to /testbed in every command and output.

How it was built

From affine-dedup (Dendrite):

1. Task-level dedup: tasks matching one of our benchmarks were dropped. Runs on tasks albedo already has were kept, under albedo's own task id when that id points at the same checkout the run had.
2. Trajectory-level dedup: no run is identical to another or to any run albedo already has.
3. Conversion: every tool call became one bash command with what that command prints; edits became a python3 heredoc that replaces the exact text; harness wrappers, to-do and memory tools were removed.
4. Solution leaks dropped: runs that got the fix from outside their own work (SWE-smith's Bug Patch history, the task's own PR in git log --all, fetching upstream code over the network).
5. Benchmark leaks dropped with albedo's scripts/benchmark_leaks.py.
6. Validated against mini-coder's format with albedo's own format checks.

Across all seven datasets: 10,456 runs exported, 9,958 converted, 2,288 dropped as solution leaks, 113 as benchmark leaks, 7,557 published.

Contents

Resolved	Rows
unknown	944
not resolved	123
resolved	83

Harness	Rows
mini_swe_textbased	1,150

Task pool	Rows
swelego	431
scaleswe	415
swerebench_v2	124
multiswe	114
r2e_gym	66

Language	Rows
python	931
go	86
js	39
rust	22
java	17
c	13
php	12
ts	12
julia	9
kotlin	3
dart	2
csharp	2
lua	1
scala	1

Columns

Column	Meaning
instance_id	task id, in the style albedo's repo-context service resolves to a repository and commit
messages	the trajectory: system prompt, task, then assistant turns (THOUGHT + one ```bash command) and <returncode>/<output> observations, mini-coder's format
family	bug family: pr for real pull requests; SWE-smith's lm, combine, mechanical otherwise
repo	owner__repo
language	repository language
verified	true when the run is known to have resolved the task
resolved	resolved, not resolved or unknown (Affine's own verdict where it gave one)
patch	the task's reference patch where the upstream record has one, else empty
model	model that produced the run
harness	agent harness the run was recorded in
id_kind	where instance_id comes from: albedo's own id for the task, Open-SWE's, or the task pool's
upstream_id	the task id in the upstream record
affine_source	task pool (swesmith, scaleswe, multiswe, swerebench_v2, r2e_gym, swelego, or SWE-Lego)
origin	Affine corpus or SWE-Lego
source_uid	where the raw run is: Affine chunk URL and line, or SWE-Lego file and row
submitted	whether the run ends with the submit command (false: cut by the step limit)
affine_fold_group	Affine's reason for keeping the run (completion, king_fail, ...), when it gave one

License

Not settled. Affine publishes no license for its corpus (none on data.affine.io or in AffineFoundation/affine). The task records come from public datasets under their own terms.