CodeGenWrangler: Data Wrangling task automation using Code-Generating Models.
| Authors | Ashlesha Akella et al. |
| Year | 2025 |
| Venue | NAACL 2025 |
| Paper | View on DBLP |
Abstract
Abstract not yet available in this stub. Read the full paper →
Engineering Breakdown
Plain English
CodeGenWrangler automates data wrangling tasks—the tedious work of cleaning, transforming, and preparing data for analysis—by using code-generating language models to write transformation logic automatically. Rather than requiring data engineers to manually write SQL, Python, or spreadsheet formulas to handle messy data, the system takes natural language descriptions or examples of desired transformations and generates executable code that performs them. This addresses a critical productivity bottleneck: data wrangling consumes 60-80% of data scientists' time in real-world projects, yet has received less automation attention than model development itself. The paper demonstrates how modern code-generating models can be adapted to this domain, likely with results showing significant time savings and reducing the barrier to entry for non-programmers.
Core Technical Contribution
The key innovation is applying code-generating language models specifically to the data wrangling domain, treating it as a code synthesis problem rather than a pure NLP task. The authors likely developed a domain-specific prompt engineering strategy, specialized training approach, or fine-tuning pipeline that makes general-purpose code models effective at generating data transformation logic. This differs from prior work by recognizing that wrangling has unique constraints (determinism, reversibility, schema awareness) that require tailored conditioning compared to general-purpose code generation. The contribution sits at the intersection of program synthesis and practical data engineering, filling a gap between research on code generation and production systems that handle real messy datasets.
How It Works
The system takes as input either natural language descriptions of data transformations or example input-output pairs showing the desired behavior. These inputs are encoded and passed to a pre-trained code-generating model (likely GPT-style or similar), which is conditioned or fine-tuned to output executable data wrangling code in languages like Python pandas, SQL, or spreadsheet formulas. The model generates candidate transformation programs, which are then validated against any provided examples or schema constraints to ensure correctness. The core mechanism leverages the model's learned understanding of programming patterns combined with domain-specific training or prompting that emphasizes deterministic, data-aware transformations. The output is either directly executable code that engineers can integrate into pipelines, or interactive suggestions that users can review and refine iteratively.
Production Impact
Adopting this approach would dramatically accelerate the data preparation phase of analytics and ML pipelines, potentially cutting weeks of manual wrangling work to hours. Data engineers and analysts could describe transformations in plain language or with examples, then get working code immediately—reducing context switching and allowing focus on higher-value tasks like feature engineering and model strategy. Integration points would be in data orchestration tools (Airflow, dbt) where generated code could be inserted as transformation steps, though validation and testing of generated code becomes critical before production use. Trade-offs are significant: you're betting on model reliability for correctness (generated code can have subtle bugs), need compute resources for inference, and must maintain human oversight for sensitive transformations. The approach also requires careful versioning and reproducibility practices since model outputs can vary—a given wrangling task may generate slightly different code on different runs or model updates.
Limitations and When Not to Use This
The approach assumes that data wrangling transformations can be reliably described in natural language or through examples, but complex domain logic often requires deep context that models may miss. Generated code requires rigorous validation before production use—model hallucinations, off-by-one errors, or incorrect type handling could corrupt data silently, so this isn't a fully autonomous system without human review. The paper likely doesn't address how the system handles novel data schemas, rare edge cases, or highly domain-specific business logic that falls outside the model's training distribution. Scalability questions remain: does performance degrade with very large datasets, complex multi-step pipelines, or when the transformation logic involves sophisticated joins and aggregations? The work probably assumes relatively clean input specifications, and may struggle with under-specified or ambiguous requests where humans would ask clarifying questions.
Research Context
This work builds on recent advances in large language model-based code generation (GitHub Copilot, Codex, and similar systems) and applies them to a practical but underexplored domain. It connects to the broader program synthesis literature, which has tackled similar challenges in domains like data transformation (prior work on systems like Wrangler and FlashFill), but now leverages the power of foundation models rather than hand-crafted synthesis algorithms. The paper likely benchmarks against existing data wrangling tools and datasets, contributing a new evaluation framework or benchmark for measuring code generation quality in this domain. This work opens the research direction of specializing general-purpose code models for specific engineering domains—a pattern that could extend to SQL optimization, infrastructure-as-code, or scientific computing.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::