Revisiting SWE-Bench: On the Importance of Data Quality for LLM-Based Code Models
Bibliographic record
Abstract
The use of Large Language Models (LLMs) for code generation has emerged as a rapidly growing field, gaining substantial traction within software engineering. However, ensuring the reliability and accuracy of generated code requires robust evaluation frameworks. To address this gap, Carlos et al. introduced the SWE-bench dataset, which consists of 2,294 GitHub issues paired with their corresponding pull requests, collected from 12 prominent Python repositories. This dataset has become a key benchmark for evaluating code generation models, with resolution rates prominently featured on the SWE-bench leaderboard. Despite its widespread adoption, the dataset has yet to undergo a systematic reliability assessment. Motivated by this gap, we conducted the first empirical study aimed at evaluating the reliability of the SWE-Bench dataset to ensure it provides meaningful and realistic model evaluations. We centered our analysis on the highest-performing model reported on the leaderboard at the time of the study: SWE-Agent + GPT-4. A thorough investigation was conducted by comparing the model-generated patches with the corresponding pull requests from the dataset. Our findings revealed two key issues: (1) 32.67% of successful cases were influenced by solution leakage, and (2) 31.08% succeeded due to weak test cases. When these problematic instances were excluded, the resolution rate of SWE-Agent + GPT-4 dropped from 12.47% to 3.97%.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.032 | 0.222 |
| Meta-epidemiology (narrow) | 0.002 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.002 |
| Bibliometrics | 0.003 | 0.003 |
| Science and technology studies | 0.001 | 0.003 |
| Scholarly communication | 0.006 | 0.009 |
| Open science | 0.004 | 0.005 |
| Research integrity | 0.002 | 0.006 |
| Insufficient payload (model declined to judge) | 0.002 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".