File size: 8,170 Bytes
acc3214
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b0fdd49
acc3214
 
 
 
 
 
 
 
 
3eb7bef
acc3214
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3eb7bef
acc3214
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3eb7bef
acc3214
 
 
 
 
 
 
 
3eb7bef
acc3214
 
3eb7bef
acc3214
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3eb7bef
 
acc3214
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
---
language:
- tr
license: mit
pretty_name: "THD-Temporal: Turkish News Stream 1997-2019"
size_categories:
- 10K<n<100K
task_categories:
- text-classification
task_ids:
- topic-classification
tags:
- turkish
- news
- concept-drift
- temporal
- online-learning
- stream-learning
- continual-learning
configs:
- config_name: default
  data_files:
  - split: finetune
    path: data/finetune.parquet
  - split: stream
    path: data/stream.parquet
- config_name: full
  data_files:
  - split: train
    path: full/thd_temporal_full.parquet
---

# THD-Temporal: a 23-year Turkish news stream for online / continual text classification

**THD-Temporal** is a chronologically ordered stream of **78,171 Turkish news articles and columns published on
Milliyet.com between 1997 and 2019**, labelled with **10 news categories**. It is designed for studying
**concept drift, online (stream) learning and continual learning** of Turkish language models: the vocabulary and the
category distribution change strongly over the two decades (e.g. *Teknoloji-Bilim* is almost absent before 2008 and
reaches >30% of the articles in 2017–2018; the shares of *Ekonomi* and *Dünya* shrink while *Gündem* grows).

> **Credit.** THD-Temporal is a cleaned, re-documented derivative of the
> **Turkish News Corpus (Türkçe Haber Derlemi)** collected by **Ömer Gözüaçık** for the *Açık Kaynak Hackathon
> Programı 2020*: <https://github.com/ogozuacik/turkce-haber-derlemi> (MIT License, © 2020 Ömer Gözüaçık).
> All article selection and the original 12-class labelling are his work; please cite the original corpus
> (see *Citation*) whenever you use this dataset.

**Links:** code and benchmark: <https://github.com/PouyaGhahramanian/LLM-OFA-TR> · Hugging Face: <https://huggingface.co/datasets/p0uy4/thd-temporal> ·
Kaggle: <https://www.kaggle.com/datasets/pooyagh/thd-temporal>

## What THD-Temporal adds to the original corpus

| Step | Detail |
|---|---|
| Article selection & labels | Taken unchanged from the original `filtrelenmis_temizlenmis_derlem.csv.gz` (81,407 articles, 12 categories). |
| Original text restored | The filtered file is lower-cased with numbers/punctuation removed; the original title, summary and body are restored from the raw scrape (`milliyet_derlem.csv.gz`) by joining on `link` (100% match). |
| Dates repaired | In the filtered file, day and month are swapped for every date with day ≤ 12 (29,427 articles); dates are re-parsed from the raw `dd.mm.yyyy` strings. |
| 10-class label | `Ege` (regional edition), `Pazar` and `Cumartesi` (weekend supplements) are not topical and are merged into **`Diğer`**; the original 12-class `category` and the Milliyet section `category_original` are kept. |
| Quality flags | Excluded from the benchmark but kept, flagged, in the `full` config: 19 invalid dates (1899); 2,247 articles whose only date is the June/July 2020 crawl date (93% sports, mostly from the Skorer site; article ids indicate publication in roughly 2009–2019, exact dates unknown; `exclusion = out_of_range_year`, their `date` is the crawl date); 970 exact duplicate articles. |
| Benchmark split | Chronological: **`finetune`** = 1997–2007 (45,384), **`stream`** = 2008–2019 (32,787), sorted by date. |

## Label distribution

| label | finetune (1997–2007) | stream (2008–2019) | total |
|---|---:|---:|---:|
| Gündem | 9,502 | 11,515 | 21,017 |
| Siyaset | 9,325 | 5,193 | 14,518 |
| Ekonomi | 8,701 | 2,456 | 11,157 |
| Diğer | 6,010 | 2,601 | 8,611 |
| Spor | 6,335 | 2,213 | 8,548 |
| Dünya | 2,971 | 1,814 | 4,785 |
| Magazin | 1,001 | 2,779 | 3,780 |
| Teknoloji-Bilim | 54 | 3,042 | 3,096 |
| Yaşam | 770 | 832 | 1,602 |
| Kültür-Sanat | 715 | 342 | 1,057 |
| **total** | **45,384** | **32,787** | **78,171** |

Yearly volume: min 1,878 (1997), max 6,703 (2006), mean 3,399 articles. About 30%
of the items are opinion columns (`is_column`).

## Columns

| column | description |
|---|---|
| `id` | stable identifier (`thd-<row of the original filtered file>`) |
| `date`, `year` | publication date (YYYY-MM-DD, repaired) and year |
| `title`, `summary`, `news` | original headline, summary (repeats the title for 17.5% of the items: 30% of `finetune`, 0.2% of `stream`) and full text |
| `label` | 10-class target: Diğer, Dünya, Ekonomi, Gündem, Kültür-Sanat, Magazin, Siyaset, Spor, Teknoloji-Bilim, Yaşam |
| `category` | original 12-class label of the source corpus |
| `category_original` | Milliyet section / sub-category (70 values in the benchmark, 107 in the `full` config) |
| `link`, `url` | relative Milliyet path and full URL |
| `is_column`, `author` | opinion column flag and columnist slug |
| `date_repaired` | the date in the source file had day/month swapped |
| `split` | `finetune` (1997–2007) or `stream` (2008–2019) |

The `full` config additionally has `in_benchmark` and `exclusion` (`invalid_date`, `out_of_range_year`, `duplicate`).

## Usage

```python
from datasets import load_dataset

# splits: finetune (1997-2007) and stream (2008-2019), both in chronological order
ds = load_dataset("p0uy4/thd-temporal")
stream = ds["stream"].to_pandas()
```

```python
import kagglehub, pandas as pd
path = kagglehub.dataset_download("pooyagh/thd-temporal")
df = pd.read_parquet(f"{path}/thd_temporal.parquet")
```

Model input used in our experiments: `title + ". " + summary`; the summary is skipped when it repeats the title, and the
`". "` becomes a single space when the title already ends in `. ! ? … : ; " '`.

## Benchmark: on-the-fly adaptation (OFA) of Turkish LLMs

The reference implementation and results (six Turkish encoders × {Static, OFA with AdamW, OFA with Adaptimizer},
pretrained and fine-tuned settings, prequential evaluation) are in <https://github.com/PouyaGhahramanian/LLM-OFA-TR>. The protocol follows
LLM-OFA (Ghahramanian et al., CIKM 2025), which introduced the English counterpart 1M-News.

## Limitations

* Single source (Milliyet.com); labels are newspaper sections, so some classes (Gündem, Diğer) are broad.
* Strong class imbalance and abrupt label-distribution changes are intentional properties of the stream, but should be
  kept in mind when comparing accuracy and macro-F1.
* Article texts belong to their publisher; as in the original corpus they are shared for research purposes.

## Citation

Please cite the original corpus and, if you use the benchmark protocol, LLM-OFA:

```bibtex
@misc{gozuacik2020turkcehaberderlemi,
  author       = {G{\"o}z{\"u}a{\c{c}}{\i}k, {\"O}mer},
  title        = {T{\"u}rk{\c{c}}e Haber Derlemi ve Haberlerin Kategorik S{\i}n{\i}fland{\i}r{\i}lmas{\i}},
  year         = {2020},
  howpublished = {\url{https://github.com/ogozuacik/turkce-haber-derlemi}}
}
@inproceedings{ghahramanian2025ofa,
  author    = {Pouya Ghahramanian and Sepehr Bakhshi and Fazli Can},
  title     = {LLM-OFA: On-the-Fly Adaptation of Large Language Models to Address Temporal Drift Across Two Decades of News},
  booktitle = {Proceedings of the 34th ACM International Conference on Information and Knowledge Management (CIKM '25)},
  year      = {2025},
  doi       = {10.1145/3746252.3760846}
}
```

## Acknowledgements

Prepared within TÜBİTAK 1001 project no. **125E060**. We thank Ömer Gözüaçık for collecting and openly sharing the
original corpus.

## Türkçe özet

THD-Temporal, Ömer Gözüaçık tarafından derlenip açık kaynak olarak paylaşılan **Türkçe Haber Derlemi**'nden
(Milliyet.com, 1997–2019) türetilmiş, zaman sırasına dizilmiş 78,171 haberlik bir akıştır. Orijinal derlemdeki
haber seçimi ve etiketler korunmuş; haberlerin özgün (büyük-küçük harf ve noktalama içeren) metinleri geri
kazandırılmış, gün/ay karışıklığı içeren tarihler düzeltilmiş, bölgesel/hafta sonu ekleri (Ege, Pazar, Cumartesi)
"Diğer" sınıfında birleştirilerek 10 sınıflı bir etiket oluşturulmuş ve 1997–2007 (ince-ayar) / 2008–2019 (akış)
şeklinde zamansal bir bölme tanımlanmıştır. Veri kümesi, Türkçe dil modellerinin kavram kayması altında çevrimiçi
(akış) ve sürekli öğrenme başarımlarının incelenmesi için hazırlanmıştır.