README.md

metadata

language:
  - en
license: apache-2.0
task_categories:
  - text-classification
tags:
  - license-detection
  - fossology
  - nirjas
  - binary-classification
  - code
  - legal
pretty_name: Nirjas Gate Dataset
size_categories:
  - 10K<n<100K

Nirjas Gate Dataset

Binary classification dataset for the Nirjas gate — a recall-first classifier that answers: "Is this text actual license text worth routing to Atarashi for identification?"

This is a narrow gate, not a broad "license-related" detector. Its job is to route actual license bodies (MIT text, GPL headers, Apache notices) to Atarashi while correctly rejecting license references ("see LICENSE file"), copyright notices, code comments, and package metadata.

Dataset Summary

split	total	license_related	not_license_related
train	55,610	30,232	25,378
validation	6,877	3,707	3,170
test	6,895	3,770	3,125

Splits are content-hash frozen — each text always lands in the same split regardless of dataset rebuilds, so benchmark deltas are real and duplicates cannot leak across train/test.

Dataset Structure

Columns

• text — the text fragment (license body, code comment, legal notice, etc.)
• label — 0 = license_related, 1 = not_license_related (ClassLabel)
• source — origin: scancode, fossology, generated, synthetic, shortform, augmented, code_corpus
• negative_type — for negatives: generic_code_comment, same_register_notice, license_discussion, copyright_discussion, todo_fixme, commented_code

Label Convention

label	int	meaning
license_related	0	actual license text — route to Atarashi
not_license_related	1	everything else — discard

Positive sources (train)

source	count
scancode (ScanCode LicenseDB)	18,931
shortform (SPDX identifiers + snippets)	4,820
synthetic/rare (LLM-augmented)	4,262
fossology (licenseRef.json)	1,681
augmented	538

Negative types (train)

type	count	description
generic_code_comment	6,324	real source comments from the-stack-smol
same_register_notice	5,177	boilerplate legal-register non-licenses
todo_fixme	4,264	TODO/FIXME/HACK comments
license_discussion	4,232	commentary about licenses, not license text
copyright_discussion	2,769	copyright attribution, not license grants
commented_code	2,612	commented-out source code

Usage

from datasets import load_dataset

ds = load_dataset("rycerzes/nirjas-dataset")
train = ds["train"]

Deployed Gate

The deployed model is potion-base-32M (StaticModelForClassification) at threshold 0.20 (recall-first, NOT argmax). At this threshold:

metric	synthetic test set	real corpus (the-stack-smol + ScanCode, n=600)
License recall	0.9952	1.0000
FPR	0.0227	0.0133

Generation Pipeline

Built by minerva-dataset-pipeline:

1. Fetch license texts from ScanCode LicenseDB + FOSSology licenseRef.json
2. Sliding-window split with junk-fragment filtering
3. LLM-assisted augmentation for rare licenses
4. Hard negative generation (LLM + template-based)
5. Content-hash frozen train/val/test split
6. Class balancing

Downstream Task

Part of the FOSSology Nirjas ML pipeline:

source file → Nirjas gate → [license_related] → Atarashi (which license?)