csv-nose
A Rust port of the Table Uniformity Method for CSV dialect detection.
Background
This crate implements the algorithm from "Detecting CSV File Dialects by Table Uniformity Measurement and Data Type Inference"[^1] by W. García.
Benchmarked against the standard CSVsniffer test suites, csv-nose is the most accurate and robust of the dialect sniffers compared; it leads outright on every dataset on both success ratio and F1; is the only tool to have 100% quote detection accuracy on the W3C-CSVW and POLLOCK test suites; and the only tool that never errors on a single file.[^bench]
This implementation of the Table Uniformity Method achieves 99.55%[^3] accuracy on the W3C-CSVW test suite by:
- Testing multiple potential dialects (delimiter × quote × line terminator combinations)
- Scoring each dialect based on table uniformity (consistent field counts)
- Scoring based on type detection (consistent data types within columns)
- Selecting the dialect with the highest combined gamma score
[^1]: García W. Detecting CSV file dialects by table uniformity measurement and data type inference. Data Science. 2024;7(2):55-72. doi:10.3233/DS-240062
[^bench]: Across the five CSVsniffer benchmark suites — POLLOCK, W3C-CSVW, and the CSV Wrangling set with its CODEC and MESSY subsets — csv-nose ranks first among the six tools compared on both success ratio and F1 on every suite, and is the only tool with a 0% error rate on every suite. As of v1.2.0 this includes the filtered CODEC and MESSY subsets, where DuckDB's sniff_csv previously tied csv-nose on CODEC and narrowly led on MESSY (while placing last on POLLOCK); both now trail. All F1 figures are derived with a single consistent methodology from each tool's success and error ratios. See the Accuracy Benchmarks tables for full per-tool figures.
Installation
As a library
As a CLI tool
With HTTP support (for remote URLs)
Library Usage
use ;
let mut sniffer = new;
sniffer.sample_size;
let metadata = sniffer.sniff_path.unwrap;
println!;
println!;
println!;
println!;
println!;
CLI Usage
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
) ()
) ()
)
)
)
)
)
)
)
)
)
)
)
)
)
Remote URL Support
When built with the http feature, csv-nose can sniff remote CSV files directly from URLs:
# Build with HTTP support
# Sniff remote CSV
# Limit bytes fetched (useful for large remote files)
The HTTP feature uses Range requests when supported by the server to minimize data transfer. If the server doesn't support Range requests, it falls back to downloading and truncating at the sample size limit.
API Compatibility
This library is designed as a drop-in replacement for qsv-sniffer used by qsv. The public API mirrors qsv-sniffer for easy migration:
use ;
let mut sniffer = new;
sniffer
.sample_size
.date_preference
.delimiter
.quote;
Accuracy Benchmarks
csv-nose is benchmarked against the same test datasets used by CSVsniffer, enabling direct accuracy comparison with other CSV dialect detection tools like python v3's csv.Sniffer, CleverCSV and DuckDB's sniff_csv.
Success Ratio
The table below shows the dialect detection success ratio. Accuracy is measured using only files that do not produce errors during dialect inference.
| Data set | csv-nose |
CSVsniffer MADSE[^2] | CSVsniffer |
CleverCSV |
csv.Sniffer |
DuckDB sniff_csv |
|---|---|---|---|---|---|---|
| POLLOCK | 98.65% | 95.27% | 96.55% | 95.17% | 96.35% | 84.14% |
| W3C-CSVW[^3] | 99.55% | 94.52% | 95.39% | 61.11% | 97.69% | 99.08% |
| CSV Wrangling | 94.97% | 90.50% | 89.94% | 87.99% | 84.26% | 91.62% |
| CSV Wrangling CODEC | 94.37% | 90.14% | 90.14% | 89.44% | 84.18% | 92.25% |
| CSV Wrangling MESSY | 93.65% | 89.60% | 89.60% | 89.60% | 83.06% | 91.94% |
[^2]: CSVsniffer MADSE is a variant of CSVsniffer using the Mean Absolute Deviation and Shannon Entropy (MADSE) method for dialect detection. [^3]: csv-nose is optimized for the W3C CSV on the Web Test Suite - reaching 99.55% accuracy.
Failure Ratio
The table below shows the failure ratio (errors during dialect detection) for each tool.
Note: "Errors" are files that caused crashes or exceptions during processing (e.g., encoding issues, malformed data). This is distinct from "failures" where a file was successfully processed but the wrong dialect was detected. A 0% error rate means all files were processed without crashes, even if some detections were incorrect.
| Data set | csv-nose |
CSVsniffer MADSE |
CSVsniffer |
CleverCSV |
csv.Sniffer |
DuckDB sniff_csv |
|---|---|---|---|---|---|---|
| POLLOCK [148 files] | 0.00% | 0.00% | 2.03% | 2.03% | 7.43% | 2.03% |
| W3C-CSVW [221 files] | 0.00% | 0.91% | 1.81% | 2.26% | 41.18% | 1.81% |
| CSV Wrangling [179 files] | 0.00% | 0.00% | 0.56% | 0.56% | 39.66% | 0.00% |
| CSV Wrangling CODEC [142 files] | 0.00% | 0.00% | 0.00% | 0.00% | 38.03% | 0.00% |
| CSV Wrangling MESSY [126 files] | 0.00% | 0.79% | 0.79% | 0.79% | 42.06% | 0.79% |
F1 Score
The F1 score is the harmonic mean of precision and recall. To compare every tool on equal footing, F1 is derived consistently from the success and error ratios above[^4]:
- precision = correct detections ÷ files processed without error
- recall = correct detections ÷ all files (files that error count as missed detections)
This penalizes a tool for both wrong detections and processing errors, so F1 equals the success ratio for any tool with a 0% error rate (e.g. csv-nose on every dataset).
| Data set | csv-nose |
CSVsniffer MADSE |
CSVsniffer |
CleverCSV |
csv.Sniffer |
DuckDB sniff_csv |
|---|---|---|---|---|---|---|
| POLLOCK | 0.986 | 0.953 | 0.956 | 0.942 | 0.926 | 0.833 |
| W3C-CSVW | 0.995 | 0.941 | 0.945 | 0.604 | 0.724 | 0.982 |
| CSV Wrangling | 0.950 | 0.905 | 0.897 | 0.877 | 0.634 | 0.916 |
| CSV Wrangling CODEC | 0.944 | 0.901 | 0.901 | 0.894 | 0.644 | 0.923 |
| CSV Wrangling MESSY | 0.937 | 0.892 | 0.892 | 0.892 | 0.609 | 0.916 |
[^4]: F1 = 2·P·R/(P+R) with the precision/recall definitions above, i.e. F1 = 2·s·(1−e)/(2−e) where s is the success ratio and e the error ratio. Competitor F1 values are derived from their reported success/error ratios (sourced from the CSVsniffer benchmark) using this same formula, rather than reproduced from a separate measurement, so all columns are directly comparable.
Component Accuracy
csv-nose's delimiter and quote detection accuracy on each dataset:
| Data set | Delimiter Accuracy | Quote Accuracy |
|---|---|---|
| POLLOCK | 98.65% | 100.00% |
| W3C-CSVW | 99.55% | 100.00% |
| CSV Wrangling | 95.53% | 99.44% |
| CSV Wrangling CODEC | 95.07% | 99.30% |
| CSV Wrangling MESSY | 94.44% | 99.21% |
NOTE: See ACCURACY.md for details on accuracy breakdowns and known limitations.
Benchmark Setup
The benchmark test files are not included in this repository. To run benchmarks, first clone CSVsniffer and copy the test files:
# Clone CSVsniffer (if not already available)
# Copy test files to csv-nose
Running Benchmarks
Once the test files are in place:
# Run benchmark on POLLOCK dataset
# Run benchmark on W3C-CSVW dataset
# Run benchmark on CSV Wrangling dataset (all 179 files)
# Run benchmark on CSV Wrangling filtered CODEC (142 files)
# Run benchmark on CSV Wrangling MESSY (126 non-normal files)
# Run integration tests with detailed output.
# --include-ignored is required: these tests are #[ignore] by default so that
# `cargo test` on a fresh clone (which has no benchmark corpora) stays green.
License
MIT license (LICENSE-MIT or https://opensource.org/licenses/MIT)
Naming
The name "csv-nose" is a play on words, combining "CSV" (Comma-Separated Values) with "nose," suggesting the tool's ability to "sniff out" the correct CSV dialect. "Nose" also sounds like "knows," implying expertise in CSV dialect detection.
AI Contributions
Claude Code using Opus 4.5 was used to assist in code generation and documentation. All AI-generated content has been reviewed and edited by human contributors to ensure accuracy and quality.