EdwardoSunny/swe-rubrics-100

SWE code-quality rubrics — 227 distinct criteria (target 100)

Falsifiable criteria about why a code change fails its tests, mined from agent attempts on real repositories. Each rubric makes one checkable prediction — a named unit test will fail, or a named exception will be raised — so it can be scored against an actual run instead of judged on taste.

This repo is one of three independent libraries collected to different sizes, for studying how a rubric database behaves with scale:

repo	distinct criteria
swe-rubrics-10	15
swe-rubrics-100	227
swe-rubrics-1000	1006

They are not nested — each stage shuffled the task pool with its own seed, so the small library is not a subset of the large one.

Quick start (retrieval)

from datasets import load_dataset
from sentence_transformers import SentenceTransformer
import numpy as np

ds = load_dataset("EdwardoSunny/swe-rubrics-100", split="train")
m  = SentenceTransformer("all-MiniLM-L6-v2")

# embed the field that exists for exactly this purpose
index = m.encode(ds["retrieval_text"], normalize_embeddings=True)

def top_k(issue_text, k=8):
    q = m.encode([issue_text], normalize_embeddings=True)[0]
    for i in np.argsort(-(index @ q))[:k]:
        yield ds[int(i)]

# and paste the pre-formatted block straight into a prompt
context = "\n".join(r["prompt_block"] for r in top_k(my_issue))

Two fields exist so you never have to reinvent them:

• retrieval_text — embed or index this. Title + prediction only; the "input condition" / "code requirement" scaffolding is for judging whether a criterion applies and only dilutes the retrieval signal.
• prompt_block — paste this. Already formatted for a model's context.

Contents

• 227 distinct criteria over 70 source repositories
• families: {'perf_spec': 221, 'error_focused': 6}
• mined from: {'swe-smith': 199, 'swe-gym': 28}

"Distinct" is post-deduplication: criteria were merged across tasks by an LLM adjudicator (with TF-IDF/embedding recall), so this is a count of surviving criteria rather than of curator calls.

Schema

field	meaning
rubric_id	stable id assigned at curation
criterion_sha256	hash of the prediction clause; survives rewording
retrieval_text	EMBED THIS. title + prediction, nothing else
prompt_block	PASTE THIS. pre-formatted for a model's context
family	perf_spec
predicate	what the rubric asserts, in words
title	one-line statement of the defect
input_condition	when the criterion applies
code_requirement	the ground truth it is checked against
prediction	the falsifiable claim
target_failing_test	test predicted red (perf_spec)
target_exception_class	exception predicted (error_focused)
source_instance_id	instance it was mined from
source_repo	upstream repo of that instance
source_dataset	swe-gym
library	swe_10
library_target	the K this library was collected to
text	the full rubric markdown
source_file	original filename

The criterion is stored split into its three clauses because they do different jobs: input_condition scopes when it applies, code_requirement is the ground truth it is checked against, and prediction is the falsifiable claim. Scoring a rubric means checking the prediction; the other two make that check well-defined.

Provenance and the leakage guarantee

Mined only from SWE-Gym and SWE-smith. SWE-bench was never mined — it is the held-out evaluation target.

Every row carries source_instance_id and source_repo, so the guarantee is checkable from the data rather than taken on trust: no row's source_repo appears in any of SWE-bench's 18 evaluation repos. Audited at export with 0 hits and 0 untraceable rows.

Held out: astropy/astropy, django/django, marshmallow-code/marshmallow, matplotlib/matplotlib, mwaskom/seaborn, pallets/flask, psf/requests, pvlib/pvlib-python, pydata/xarray, pydicom/pydicom, pylint-dev/astroid, pylint-dev/pylint, pytest-dev/pytest, pyvista/pyvista, scikit-learn/scikit-learn, sphinx-doc/sphinx, sqlfluff/sqlfluff, sympy/sympy

Families

• perf_spec — the code runs and is wrong; predicts a failing unit test. This dominates, because SWE work is mostly "runs and gives the wrong answer" rather than "crashes".
• error_focused — the change makes the module unimportable or otherwise raises; predicts an exception class.
• error_grader — the grader rejects the output.

Limitations

• Mined with a small tool-loop agent, so the failure modes reflect what that agent gets wrong — not the whole space of code defects.
• error_focused is a small minority. Crashes are genuinely rare in this domain; the family is thin by nature, not by neglect.
• No precision/recall is published here. Distinctness and provenance are measured; predictive quality is not, and should not be assumed.