Generating Data Engineering Code Using Llms
Bibliographic record
Abstract
Data engineering is a complex and time-consuming part of data science, critical for transforming raw data into actionable insights. This complexity stems from diverse and large data sources, data-dependent logic, and exploratory workflows. We perform an empirical evaluation of the performance of three LLMs (GPT-4o-mini, Claude-3.5-Haiku, and Gemini-2.0-Flash), which have been shown to be effective at code generation, on multi-step data engineering notebooks. Our study considers the impact of prompting with previous execution context, output samples, and the application of iterative refinement on entire notebooks and their individual steps. We benchmark performance against the ARCADE dataset and introduce a new benchmark derived from Spider 2.0 (Spider2-intents) to mitigate potential data leakage. Our results show that LLMs generate syntactically and semantically correct code, with output data match scores reaching up to 80%, and BLEU scores of 0.35 on our newly created Spider2-intents benchmark. While generated code trends toward reduced runtime, memory, and CPU usage, these improvements are not statistically significant. Further analysis reveals that ambiguity in user intents is the leading cause of functional correctness issues, accounting for 40.74% of such cases. We also observe that iterative refinement shows a modest but statistically inconclusive trend toward improved output correctness, with gains of up to 2.9% after two rounds of notebook and intent refinement each.
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 | 0.002 | 0.016 |
| Meta-epidemiology (narrow) | 0.002 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.002 |
| Science and technology studies | 0.001 | 0.001 |
| Scholarly communication | 0.001 | 0.002 |
| Open science | 0.002 | 0.002 |
| Research integrity | 0.001 | 0.002 |
| Insufficient payload (model declined to judge) | 0.004 | 0.004 |
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".