CodeLanguage / README.md
Yash1005's picture
docs: dataset card with totals, labels, token buckets
6c7a1e3 verified
metadata
license: apache-2.0
task_categories:
  - text-classification
tags:
  - code
  - language-identification
  - qwen
pretty_name: Qwen Code Language-ID SFT Dataset

Accuknoxtechnologies/CodeLanguage

SFT dataset for fine-tuning a Qwen-based guard that detects which programming languages appear in a user prompt. Each row pairs a natural-language + code prompt with a JSON target enumerating the detected languages.

This release combines the previously-separate train + test CSVs into a single train split (source files: code_langid.csv, test_dataset_langid.csv).

Schema

column description
prompt user message, possibly containing one or more code snippets
target JSON: {"is_valid": bool, "category": {"<Lang>": true, ...}}
kind one of single, multi, benign

Total Records

rows single multi benign invalid (is_valid=false)
10500 7349 2101 1050 1050

Supported Labels

25 programming languages:

AWK, Bash, Batch, C, C#, C++, Dockerfile, Go, Java, JavaScript, Kotlin, Lua, Makefile, Perl, PowerShell, Python, R, Ruby, Rust, SQL, Scala, Swift, Terraform, YAML, jq

label rows containing label
Go 508
PowerShell 505
Lua 504
Terraform 502
Rust 498
AWK 496
Kotlin 496
JavaScript 495
Batch 495
C 492
C++ 491
Scala 486
Ruby 486
jq 480
R 479
Dockerfile 475
Swift 472
C# 469
Makefile 467
Bash 466
Perl 466
SQL 465
Python 462
Java 460
YAML 448

Token-wise Bucket Split

Tokenized with Qwen/Qwen2.5-0.5B (matches the training tokenizer).

0-128 129-256 257-512 513-1024 1025-2048 2049+ min mean p50 p95 max
1529 1725 2398 2759 2060 29 23 581.8 449 1361 3035

Languages (natural language of the prompts)

Per-prompt natural-language detection (English/Korean/etc.) is not computed in this card revision. Prompts are predominantly English by construction (see build_dataset.py).

Reproduction

Generated by gpu-vm-training-langid/build_dataset.py and pushed by gpu-vm-training-langid/hf_dataset_push/push_dataset.py.