Datasets:

razhan
/

code-valid

configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
dataset_info:
  features:
    - name: repo_name
      dtype: string
    - name: path
      dtype: string
    - name: copies
      dtype: string
    - name: size
      dtype: string
    - name: content
      dtype: string
    - name: license
      dtype: string
    - name: hash
      dtype: int64
    - name: line_mean
      dtype: float64
    - name: line_max
      dtype: int64
    - name: alpha_frac
      dtype: float64
    - name: autogenerated
      dtype: bool
    - name: ratio
      dtype: float64
    - name: config_test
      dtype: bool
    - name: has_no_keywords
      dtype: bool
    - name: few_assignments
      dtype: bool
  splits:
    - name: train
      num_bytes: 432490836
      num_examples: 45961
  download_size: 159490243
  dataset_size: 432490836
license: other
language:
  - code
tags:
  - code
  - python
  - filtered
size_categories:
  - 10K<n<100K

code-valid

45,961 Python files that passed a code-quality filtering pass, each kept with its source repository, file path, declared licence, and the filter values that were computed on it.

At a glance


Rows	45,961 (single `train` split)
File type	Python only — `.py` for 100% of rows
Columns	15 (see below)
Parquet on disk	159.5 MB (432.5 MB uncompressed)
Row groups	5

Columns

Column	Type	What it holds
`repo_name`	string	source repository, e.g. `pansapiens/mytardis`
`path`	string	file path in the repository, e.g. `tardis/apps/mx_views/views.py`
`content`	string	the file's full source text (mean 9,272 chars, median 3,937, max 644,045)
`license`	string	SPDX-style licence string of the repository
`size`	string	file size in bytes, as a string (1,024 – 895,755; mean 9,231)
`copies`	string	in how many repositories the same content was found (1 for 44,880 rows, up to 11)
`hash`	int64	content hash used for the duplicate check
`line_mean`	float64	mean line length (6.5 – 99.9, mean 33.8)
`line_max`	int64	longest line
`alpha_frac`	float64	fraction of alphabetic characters (0.25 – 0.96, mean 0.60)
`ratio`	float64	filter ratio value (1.50 – 13.56, mean 3.70)
`autogenerated`	bool	`False` in every row of this snapshot
`config_test`	bool	`True` in 5,104 rows (11%)
`has_no_keywords`	bool	`True` in 301 rows
`few_assignments`	bool	`False` in every row of this snapshot

Licence distribution

Counted over all 45,961 rows — 15 distinct values, no nulls:

Licence	Rows
`mit`	14,285
`apache-2.0`	8,216
`gpl-3.0`	8,188
`gpl-2.0`	4,708
`bsd-3-clause`	4,650
`agpl-3.0`	2,307
`bsd-2-clause`	1,224
`lgpl-3.0`	806
`lgpl-2.1`	580
`unlicense`	328
`mpl-2.0`	273
`isc`	162
`cc0-1.0`	135
`epl-1.0`	77
`artistic-2.0`	22

Usage

from datasets import load_dataset

ds = load_dataset("razhan/code-valid", split="train")
row = ds[0]
print(row["repo_name"], row["path"], row["license"])
print(row["content"][:300])

If you need only permissively licensed code, filter on license in {"mit", "apache-2.0", "bsd-2-clause", "bsd-3-clause", "isc", "unlicense", "cc0-1.0"} — that keeps exactly 29,000 rows. Copyleft licences account for the remaining 16,589.

Known limitations

Licence is per repository, not per file. Rows inherit the repository's declared licence; verify before redistribution, and note that 16,589 rows carry copyleft licences (gpl-*, agpl-3.0, lgpl-*).
Duplicate content is not removed: copies > 1 for 1,081 rows, and those rows appear once per repository.
The flag columns record a single filtering pass from 2022; autogenerated and few_assignments are constant False here, so they carry no signal in this snapshot.
No language/version metadata beyond "it parses as Python", and no repository snapshot date — code drift between the crawl and today is not tracked.
Some files are minified, generated, or vendor dumps that survived the filters applied at the time.