lonewolflab/Morpheus-TR-corpus

Morpheus-TR Corpus (partial release)

A partial release of the monolingual Turkish corpus behind Morpheus and the paper Morpheus: A Morphology-Aware Neural Tokenizer and Word Embedder for Turkish. It mixes informal, academic, journalistic, and encyclopedic Turkish so that a tokenizer or embedding model sees the full range of Turkish suffixation, not just one register.

File	Lines	Words	Characters	Size
train.txt	1,679,159	55.4M	422.4M	461 MB
test.txt	186,574	6.1M	46.8M	51 MB
Total	1,865,733	61.6M	469.2M	512 MB

Every line is unique across both files: no line repeats within a split, and no test.txt line occurs in train.txt.

About this release

The full corpus used to train and evaluate Morpheus (~10 GB of raw text) was collected, cleaned, filtered, and deduplicated over a long period. It is not distributed publicly:

This repository contains about 0.5 GB drawn from the same sources and processed with the same pipeline. It is meant for three things:

• inspecting the data format, the registers, and the cleaning;
• running the TurkishMorpheus pipeline end to end;
• small-scale Turkish tokenization and language-modeling experiments.

Results computed on this subset are not the numbers reported in the paper and will differ from them.

Access to the full corpus. For research collaboration or licensing, contact the author by opening a discussion on this repository or through Tolga Şakar's Hugging Face profile.

Sources

Source	Register	What it adds
Ekşisözlük	Informal / colloquial	Spoken-language suffixation (-ymiş, -sin), idioms, slang
Dergipark	Academic / formal	Terminology and derivational morphology
Turkish news sites	Standard / journalistic	Neutral register, broad vocabulary
Turkish Wikipedia	Encyclopedic	Wide vocabulary and word-form diversity

The web registers were collected and cleaned with CorpusCollector, which documents the source URLs, scraping protocol, rate limiting, per-source extraction (HTML and URL stripping, Unicode normalization), and deduplication.

The Wikipedia portion comes from wikimedia/wikipedia, snapshot 20231101.tr, filtered line by line with wikipedia_ingest.py. A line is kept only if it:

• is 30–600 characters long after removing reference brackets ([1], [düzenle], [kaynak belirtilmeli]) and collapsing whitespace;
• contains no markup or links (http, www., [[ ]], {{ }}, |, ==, HTML tags and entities);
• has at least 3 tokens and no token longer than 34 characters;
• has ≥ 92% of its letters in the Turkish alphabet, ≤ 2% of q/w/x, and ≤ 20% non-letter, non-space characters;
• contains at least one common Turkish function word (ve, bir, bu, için, …).

Exact duplicates are then removed by hash.

Format and splits

• Plain UTF-8 text, one paragraph (occasionally a whole document) per line, no trailing newline. Lines have a median length of 193 characters and a 99th percentile of about 600; a few author-corpus lines are much longer (up to 228K characters). Original casing, punctuation, digits, and symbols are kept. The Turkish-aware case folding described in the paper happens inside the Morpheus pipeline, not in these files.
• The splits are made by analyzer.py, which runs four steps on the merged and deduplicated corpus:
  1. Strip every line and drop empty lines.
  2. Shuffle at the line level with random.seed(42).
  3. Cut at train_split_ratio = 0.90, giving a 90/10 split.
  4. Write both files.
• Because the shuffle is line-level, consecutive lines are usually unrelated: they do not form documents.
• Before the split, the merged corpus is deduplicated by exact match on the stripped line, keeping the first occurrence. This guarantees that the two splits are disjoint.

How the pipeline uses the splits

The TurkishMorpheus pipeline assigns the two files the same roles they had in the full corpus:

Use	Data
Morpheus vocabulary, all subword baselines (BPE, Unigram, byte-level BPE, WordPiece), Morfessor baseline	train.txt
Language-modeling probe: training	first 1M lines of train.txt
Language-modeling probe: validation (all intermediate evaluations)	the next 20K lines of train.txt
Language-modeling probe: test, scored once on final weights	first 100K lines of test.txt
Raw-text roundtrip evaluation	first 5,000 lines of test.txt

Usage

from datasets import load_dataset

ds = load_dataset("lonewolflab/Morpheus-TR-corpus")
print(ds)                       # DatasetDict with "train" and "test", one "text" field per line
print(ds["test"][0]["text"])

To run TurkishMorpheus on this subset, download the raw files to the path the pipeline reads:

from huggingface_hub import hf_hub_download

for name in ["train.txt", "test.txt"]:
    hf_hub_download("lonewolflab/Morpheus-TR-corpus", name, repo_type="dataset",
                    local_dir="src/model_development/artifacts/datasets/splits")

Limitations and risks

• Web text. The Ekşisözlük and news portions contain informal, opinionated, and sometimes offensive language. They were not filtered for toxicity.
• Personal information. User-generated text can contain names, usernames, or other personal details. No systematic PII removal was applied beyond URL stripping.
• Line-level shuffling. The shuffle removes document context, so the corpus is not suitable for training or evaluating long-range, document-level models.

Citation

@misc{sakar2026morpheus,
  title  = {Morpheus: A Morphology-Aware Neural Tokenizer and Word Embedder for Turkish},
  author = {Şakar, Tolga},
  year   = {2026},
  note   = {arxiv.org/abs/2606.18717}
}

Author: Tolga Şakar · Code: TurkishMorpheus · Model: Morpheus-TR-50K