nitinmurali21/prcurator-v1

PRcurator v1

A quality-graded dataset of 3,491 merged GitHub Pull Requests, scored across seven quality dimensions designed for training and evaluating code agents.

What this is

Frontier coding models are now bottlenecked on data quality, not quantity. PR threads are uniquely valuable training material because they carry the human-review signal — someone wrote code, someone else pointed out problems, the author fixed them. PRcurator scores that signal explicitly so downstream training can weight, filter, or partition by quality.

The seven dimensions

#	Dimension	Direction	Composite weight
1	change_focus (smaller, more focused diffs score higher)	+	0.15
2	human_reviewer_engagement (log-scaled comment count + thread depth)	+	0.25
3	bot_reviewer_engagement (CI / Copilot-reviewer noise)	–	0.05
4	test_inclusion (1.0 if any test file modified, 0.3 floor)	+	0.15
5	ai_authored_probability (heuristic + Haiku 4.5 judge)	–	0.30
6	comment_hygiene (per-comment off-topic detection, Haiku 4.5)	+	0.10
7	license_compatibility	gate	hard filter

The composite is a weighted blend with two inverted dimensions (bot_reviewer_engagement and ai_authored_probability enter as 1 − score). License-incompatible rows are dropped before scoring is reported.

Composition

Repositories pulled (all permissively licensed):

Role	Repo	License	PRs
anchor	microsoft/vscode	MIT	~3,000
slice	django/django	BSD-3-Clause	100
slice	pandas-dev/pandas	BSD-3-Clause	100
slice	facebook/react	MIT	100
slice	evanw/esbuild	MIT	100
slice	BurntSushi/ripgrep	MIT	100

All PRs are the most recent merged-and-non-revert PRs at pull time.

Headline findings

• 3,491 scored PRs across 6 repositories.
• 29% of recent merged microsoft/vscode PRs are AI-authored (probability ≥ 0.5 from the heuristic-plus-judge ensemble). The biggest single signal is Co-Authored-By: Copilot in commit metadata.

Downstream validation

We tested the relationship between PRcurator quality grading and a frontier LLM's ability to predict the merged diff. 100 top-quartile and 100 bottom-quartile PRs were sampled (seed 42), the pre-change file excerpts were extracted from each unified diff, and Claude Sonnet 4.6 was asked to predict the resulting unified diff. Normalized Levenshtein distance between predicted and actual diff was measured.

Statistic	Value
Median edit distance — top quartile	0.5993 (n = 100)
Median edit distance — bottom quartile	0.5661 (n = 100)
OLS coefficient on log(diff size)	+0.1000 (p = < 0.001) ← dominant predictor
Spearman partial ρ (composite vs edit_distance, controlling for diff size + file count)	+0.165 (p = 0.0208)

Top-quartile PRs are less predictable to Claude Sonnet 4.6 after controlling for diff size, with a small but statistically significant effect. The initial hypothesis — that high-quality PRs would be more predictable — was wrong, and the failure was diagnostic. Diff completion fidelity is dominated by diff size (p < 1e-14 in the regression), and bottom-quartile PRs are dominated by AI-authored mechanical changes that frontier LLMs already produce well. Top-quartile PRs are the substantive engineering work models can't yet replicate from a title and pre-image. This is the residual signal that's worth the training-cost: bottom-quartile PRs validate what models can already do, top-quartile PRs surface what they still need to learn.

Methodological caveat. Diff completion fidelity is a proxy for training usefulness, not a direct test of it. A direct test would require fine-tuning two models on the top vs bottom quartile and comparing downstream performance — this is scoped to a future version. The Spearman partial ρ above is the strongest external claim v1 makes.

Schema

The single split is train. Columns:

Column	Type	Notes
pr_id	string	<owner>/<repo>#<number>
repo	string	<owner>/<repo>
pr_number	int64	numeric PR number
title, body	string	PR text
author_login, author_is_bot	string, bool
merged_at	string	ISO-8601
lines_added, lines_deleted, files_changed	int32
n_human_comments, n_bot_comments, n_reviews_human	int32
change_focus, human_reviewer_engagement, bot_reviewer_engagement	float32	dim scores [0, 1]
test_inclusion, ai_authored_probability, comment_hygiene	float32	dim scores [0, 1]
ai_authored_signals	string	JSON-encoded list of triggered heuristic signals or LLM reason
license_spdx	string	MIT, Apache-2.0, ...
composite_score	float32	weighted blend, range [0, 1]
quality_quartile	int8	1 (top) — 4 (bottom)
diff_text	string	unified diff, truncated to 50K chars

Loading

from datasets import load_dataset
ds = load_dataset("nitinmurali21/prcurator-v1")
df = ds["train"].to_pandas()
top_quality = df[df["quality_quartile"] == 1]

Methodology, code, reproducibility

Full pipeline, scorers, and validation script live at https://github.com/nitinm21/prcurator. The data was pulled via the GitHub REST API; scoring uses Anthropic Claude Haiku 4.5 for the two LLM-judged dimensions; validation uses Claude Sonnet 4.6.

Limitations

• Single anchor repo (microsoft/vscode) provides the majority of rows. Slice repos extend coverage across languages and review cultures but are small at 100 PRs each.
• AI-authored detection is heuristic-first; LLM judge handles ~5% of ambiguous cases. May undercount AI-authored PRs that hide their origin.
• Diff completion fidelity is a proxy for "training usefulness," not a direct test. A model that perfectly memorizes vscode PRs would score arbitrarily well — Sonnet 4.6 did not train on these particular PRs at publication time, but the proxy is imperfect.
• English-only LLM judge prompts.

License

Creative Commons Attribution 4.0 International (CC-BY 4.0) on the data. Code is MIT.

Citation

@misc{prcurator2026,
  author = {Nitin Murali},
  title  = {PRcurator: A Quality-Graded Dataset of GitHub Pull Requests for Code-Agent Training},
  year   = {2026},
  url    = {https://huggingface.co/datasets/nitinmurali21/prcurator-v1}
}