smartcore-v1-data / README.md
kdirgul's picture
dataset kartı
98e2990 verified
|
Raw
History Blame Contribute Delete
2.47 kB
---
license: other
language:
- en
- tr
tags:
- pretraining
- mamba
- smartcore-v4
- decontaminated
size_categories:
- 10B<n<100B
---
# SmartCore V4 — Pretraining Verisi (12B token, EN+TR+kod+math)
**SmartCore V4** (sıfırdan ~180M Mamba-3 SISO + GQA 5:1 hibrit, EN temel + TR ikincil)
projesinin **ön-tokenize edilmiş, dekontamine** pretraining verisi.
- **Tokenizer:** [`kdirgul/smartcore-v4`](https://huggingface.co/kdirgul/smartcore-v4) → `tokenizer/` (48K SentencePiece, byte_fallback, EN+TR).
- **Paketleme:** Her doküman `encode + EOS` → ardışık **2048 token**'lık dizilere paketlendi (doc-arası carry-over). Padding yok.
- **Format:** parquet shard'lar, şema `{input_ids: list<uint16>[2048], source: str, lang: str}`. uint16 (token id < 48000).
- **Oluşturma:** 2026-06-05, `kod/faz1_01_tokenize_shard.py` (proje deposu).
## İçerik (toplam 12.00B token, 181 shard, ~18GB)
| Kaynak (`source`) | Köken | lang | Token | Shard |
|---|---|---|---|---|
| `en_fineweb_edu` | HuggingFaceFW/fineweb-edu (sample-10BT) | en | 6.60B | 99 |
| `tr_fineweb2_hq` | epfml/FineWeb2-HQ (tur_Latn) | tr | 2.64B | 40 |
| `code_codeparrot` | codeparrot/codeparrot-clean | code | 1.56B | 24 |
| `math_openwebmath` | open-web-math/open-web-math | math | 1.20B | 18 |
Her kaynak `<source>/shard_*.parquet` + `<source>/manifest.json` altında.
## Eğitim karışımı (mixture)
Shard'lar kaynak-bazlı saklanır; oranlar **eğitimde** uygulanır:
`%55 EN / %22 TR / %13 kod / %10 math`. (Decay/anneal fazında top-quartile alt küme — bkz. proje kılavuzu.)
## Dekontaminasyon
Eval paneline karşı **13-gram örtüşmesiyle** kontamine dokümanlar **çıkarıldı** (`kod/decontam.py`).
Eval panel (9/11 set): Belebele-tr, XNLI-tr, XCOPA-tr, TR-MMLU + HellaSwag, ARC-E/C, BoolQ, MMLU.
(MKQA + PIQA gate/script nedeniyle hariç.)
| Kaynak | Atlanan doküman |
|---|---|
| EN | 8587 (%0.14 — en çok; İngilizce eval örtüşmesi) |
| TR | 101 (%0.002) |
| Kod | 22 |
| Math | 619 (%0.13) |
## Yükleme
```python
from datasets import load_dataset
ds = load_dataset("kdirgul/smartcore-v4-data", data_dir="en_fineweb_edu", split="train", streaming=True)
ex = next(iter(ds)) # {"input_ids": [2048 token], "source": "...", "lang": "..."}
```
## Lisans / kullanım notu
Türev veri; kaynak lisansları geçerli: FineWeb-Edu (ODC-By), FineWeb2-HQ, codeparrot-clean,
open-web-math. Yalnız araştırma/eğitim amaçlı. Repo **private**.