SWE-Review-Traj

8,914 decision-correct agentic code review trajectories for training open-source code review models. Each trajectory captures a full multi-turn review session where an AI agent explores a repository, traces the root cause of a bug, and produces a structured review decision — all verified against executable test suites.

Project Page | Paper | Code | Benchmark | Training Data | Claude Code Plugin

About SWE-Review

SWE-Review is a framework for closing the issue-resolution loop with agentic code review. A reviewer agent independently explores the repository, traces the root cause, and compares its own diagnosis against the submitted PR — turning one-shot patch generation into an iterative generate-review-revise loop that raises resolve rates by up to +29.4 percentage points on SWE-bench Verified.

Data Collection Pipeline

SWE-rebench (~6K instances) → Patchgen (3 models) → 14,156 candidate PRs
  → Teacher Review (GLM-5, agentic) → Decision-Correct Filtering → 8,914 trajectories

Step 1: Candidate PR Generation

Three models generate patches for ~6,000 SWE-rebench instances (no overlap with SWE-Review-Bench to prevent data leakage):

Patchgen Model	Trajectories
GLM-5-FP8	3,056
Qwen3-Coder-30B-A3B-Instruct	2,991
Qwen3-30B-A3B-Instruct-2507	2,867

Step 2: Teacher Review

Open-weight GLM-5 with thinking enabled reviews each PR agentically via OpenHands-SDK. The review prompt enforces independent reconstruction: the reviewer must trace the root cause from the issue description and repository before examining the candidate patch, preventing confirmation bias.

Step 3: Decision-Correct Filtering

Only trajectories where the reviewer's decision matches ground truth are retained:

• Approve a patch that actually resolves the issue ✓
• Request changes on a patch that does not resolve the issue ✓

Quality Validation

• Semantic quality: Two independent judges (Claude Opus 4.6, GPT-5.4) rate diagnosis quality. Mean overall score >3.0/5. Cohen's κ = 0.72 (substantial inter-judge agreement).
• Functional quality: Diagnoses carry actionable information beyond the binary decision — revision RRR rises from 3% (no review) → 8% (decision only) → 21% (teacher review) → 32% (oracle).

Training Value

Review trajectories are not only useful for training reviewers — they also improve code generation when mixed with issue-resolution data:

Training Data (Qwen3-8B)	RR (one-shot fix)	Loop RRR (self-review + revise)
Issue-resolution 2k	31.2%	31.2%
+ Review 2k	36.8% (+5.6pp)	41.8% (+10.6pp)
Issue-resolution 3k	34.0%	34.0%
+ Review 3k	37.8% (+3.8pp)	41.2% (+7.2pp)

Mixed training simultaneously improves one-shot resolve rate and enables self-contained generate-review-revise loops within a single model.

Usage

Load the Dataset

from datasets import load_dataset
import json

ds = load_dataset("SWE-Lego/SWE-Review-Traj", split="train")
print(f"Total trajectories: {len(ds)}")

# Inspect a trajectory
traj = ds[0]
print(f"Instance: {traj['pr_instance_id']}")
print(f"Decision: {traj['decision']} (confidence: {traj['confidence']:.2f})")
print(f"Turns: {traj['turns']}")
print(f"Patch resolved: {traj['patch_resolved']}")

Train with LLaMA-Factory

The messages field is in ShareGPT format — directly compatible with LLaMA-Factory:

import json

ds = load_dataset("SWE-Lego/SWE-Review-Traj", split="train")
sft_data = [json.loads(row["messages"]) for row in ds]

with open("swe_review_traj.json", "w") as f:
    json.dump(sft_data, f)

Then register in dataset_info.json and train with LLaMA-Factory. See the training guide for complete instructions.

Access Judge Scores

4,548 trajectories include dual-judge quality scores:

traj = ds[0]
if traj["judge_gpt54"]:
    scores = json.loads(traj["judge_gpt54"])
    print(f"Diagnosis accuracy: {scores['diagnosis_accuracy']}/5")
    print(f"Suggestion correctness: {scores['suggestion_correctness']}/5")
    print(f"Overall: {scores['overall_score']}/5")

Schema

Field	Type	Description
pr_instance_id	string	PR instance identifier
issue_instance_id	string	Issue instance identifier
repo	string	GitHub repository
problem_statement	string	GitHub issue description
patchgen_model	string	Model that generated the candidate PR
model_patch	string	Candidate patch (unified diff)
patch_resolved	bool	Whether the patch resolves the issue
messages	string (JSON)	Full conversation in ShareGPT format
tools	string (JSON)	Tool definitions used during review
turns	int	Number of conversation turns
teacher_model	string	Teacher model used for review
decision	string	approve or request_changes
confidence	float	Reviewer confidence (0.0–1.0)
report_json	string (JSON)	Structured review report
judge_gpt54	string (JSON)	GPT-5.4 quality scores (null if not evaluated)
judge_opus46	string (JSON)	Claude Opus 4.6 quality scores (null if not evaluated)

Judge Score Fields

Each judge JSON contains:

Field	Range	Description
diagnosis_accuracy	1–5	Correctness of root-cause identification
suggestion_correctness	1–5	Correctness of suggested fixes
grounding_quality	1–5	Whether suggestions reference specific code locations
overall_score	1–5	Overall review quality
verdict	1–5	Alignment with ground-truth resolution status

Citation

@article{wang2026swereview,
  title={SWE-Review: Closing the Loop on Issue Resolution with Agentic Code Review},
  author={Wang, Ruoyu and Chen, Jierun and Wang, Shaowei and Tao, Chaofan and Yang, Sidi and Jiang, Yuxin and Yap, Kim-Hui and Shang, Lifeng and Li, Xiaohui and Bai, Haoli},
  journal={arXiv preprint},
  year={2026}
}