Chamaka8/serendip-cpt-sinhala

Serendib LLM CPT Sinhala Corpus

A large-scale, deduplicated, quality-filtered Sinhala plain-text corpus built for Continual Pre-Training (CPT) of large language models. This dataset was used to adapt Meta-LLaMA-3-8B to the Sinhala language domain as part of the Serendib LLM Honours Degree Research Project at the University of Central Lancashire (UCLan), 2025–2026.

This is one of the largest openly published Sinhala NLP corpora available, containing 23,449,223 training documents across 11 cleaned shards totalling 21.6 GB of storage.

---

Dataset Statistics

Splits

Split	File(s)	Entries	Notes
Train (clean)	train_shards_clean/train_000–010.jsonl	23,449,223	Primary training data
Validation	val.jsonl	23,150	Full validation set
Val Small	val_small.jsonl	2,000	Quick validation subset
Train Small	train_small.jsonl	20,000	Quick training subset
Grand Total		23,494,373

Per-Shard Breakdown (Clean)

Shard	Entries	Text Size	Avg Doc Length
train_000.jsonl	2,410,010	549.0 MB	238 chars
train_001.jsonl	2,402,042	548.6 MB	239 chars
train_002.jsonl	2,401,723	549.0 MB	239 chars
train_003.jsonl	2,118,331	535.9 MB	265 chars
train_004.jsonl	348,806	425.8 MB	1,279 chars
train_005.jsonl	2,821,090	511.7 MB	190 chars
train_006.jsonl	2,805,252	511.0 MB	191 chars
train_007.jsonl	2,744,722	508.5 MB	194 chars
train_008.jsonl	2,668,897	505.2 MB	198 chars
train_009.jsonl	2,585,353	501.6 MB	203 chars
train_010.jsonl	142,997	28.1 MB	206 chars
Total	23,449,223	~5.17 GB text	~234 chars avg

Note: train_004 has a significantly higher average document length (1,279 chars) compared to other shards (~200 chars), indicating it contains long-form documents such as Wikipedia articles and government publications.

Repository Storage

Total LFS storage: 21.6 GB (raw + clean shards, scripts, validation sets)

---

Schema

Each document is stored as a JSON object. Fields vary slightly across shards due to source diversity:

Field	Type	Present In	Description
id	string	All shards	Unique document identifier
text	string	All shards	Sinhala plain text content
source	string	All shards	Origin website or corpus source
label	string	train_000, val	Document category label
lang	string	train_001–003	Language tag (si)
src	string	train_001–003	Source URL or reference

Sample Entry

{
  "id": "si_001234",
  "text": "ශ්‍රී ලංකාව දිවයිනක් වන අතර එය ඉන්දියාවේ දකුණු කෙළවරට නුදුරින් පිහිටා ඇත...",
  "source": "sinhala.wikipedia.org"
}

---

Data Sources

The corpus was assembled from six distinct Sinhala web source categories to maximise linguistic diversity and domain coverage:

Source Type	Examples	Content Style
Sinhala Wikipedia	sinhala.wikipedia.org	Encyclopaedic, formal prose
News portals	AdaDerana, NewsFirst, Lankadeepa	News articles, formal
Government publications	Official Sinhala documents	Formal, institutional
Online forums	Sinhala discussion boards	Informal, conversational
Educational materials	Sinhala-medium academic content	Educational, structured
Literary blogs	Sinhala poetry and essay sites	Creative, literary

---

Cleaning Pipeline

A 7-stage cleaning pipeline was applied to the raw shards to produce train_shards_clean/:

Step	Method	Threshold / Detail
1. HTML removal	BeautifulSoup get_text() + regex	Strips nav links, ads, JS snippets
2. Encoding filter	Repeated character detection	Entries with >5 consecutive identical chars removed
3. Numeric filter	Numeric content ratio check	>15% numeric content discarded
4. Exact deduplication	SHA-256 hash of normalised string	Identical documents removed
5. Near-dedup	5-gram Jaccard similarity	Threshold 0.85 — near-identical republications removed
6. Length filter	Token count check	<20 or >2,048 tokens discarded
7. Language filter	Unicode range U+0D80–U+0DFF	<40% Sinhala characters discarded

---

Repository Structure

serendip-cpt-sinhala/
├── data/
│   └── raw/
│       ├── train_shards/           # Raw scraped shards (000–010)
│       ├── train_shards_clean/     # Cleaned & deduplicated shards (000–010)
│       ├── train_small.jsonl       # 20,000 entry quick-train subset
│       ├── val.jsonl               # 23,150 entry validation set
│       └── val_small.jsonl         # 2,000 entry quick-val subset
├── train_scripts/
│   ├── train_cpt.py                # Continual pre-training script
│   ├── train_sft_lora.py           # SFT LoRA fine-tuning script
│   ├── train_dpo.py                # DPO alignment script
│   ├── merge_lora.py               # LoRA adapter merge script
│   ├── load_data.py                # Dataset loading utilities
│   ├── rebuild_clean_shards.py     # Cleaning pipeline script
│   ├── sample_preview.py           # Data preview utility
│   ├── sanity_check_dataset.py     # Dataset validation script
│   └── requirements.txt            # Python dependencies
├── runpod/
│   ├── run_sft_lora.sh             # RunPod training launch script
│   ├── setup_env.sh                # RunPod environment setup
│   └── .env.example                # Environment variable template
├── test_base.py                    # Base model test script
├── test_cpt.py                     # CPT model test script
├── mini_run.py                     # Quick sanity check run
└── compare_small_vs_cpt.py         # Base vs CPT comparison script

---

Usage

Load with Hugging Face Datasets

from datasets import load_dataset

# Load full training set
ds = load_dataset("Chamaka8/serendip-cpt-sinhala", split="train")
print(f"Training entries: {len(ds):,}")
print(ds[0])

# Load validation set
val = load_dataset("Chamaka8/serendip-cpt-sinhala", split="validation")
print(f"Validation entries: {len(val):,}")

Load a Single Shard

from datasets import load_dataset

shard = load_dataset(
    "json",
    data_files="hf://datasets/Chamaka8/serendip-cpt-sinhala/data/raw/train_shards_clean/train_000.jsonl"
)
print(shard)

Use for Continual Pre-Training

from datasets import load_dataset
from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("Chamaka8/serendib-tokenizer")
ds = load_dataset("Chamaka8/serendip-cpt-sinhala", split="train", streaming=True)

for example in ds:
    tokens = tokenizer(example["text"], truncation=True, max_length=2048)
    # Feed to your CPT training loop
    break

---

Trained Models

This corpus was used to train the following models:

Model	Description	Link
serendib-llm-cpt-llama3-8b	CPT base model trained on this corpus	🤗 Hub
Serendip-LLM-CPT-SFT-v2	SFT model built on CPT base	🤗 Hub

---

Related Datasets

Dataset	Description	Link
Serendip-sft-sinhala	603,000 Sinhala instruction-response pairs	🤗 Hub
SerendibLLM-PoemSong-Dataset	19,184 Sinhala poem and song generation entries	🤗 Hub

---

Hardware Used for Training

• GPU: NVIDIA RTX A4500 (20 GB VRAM) and A100 (40 GB VRAM) on RunPod
• Framework: PyTorch + HuggingFace Transformers + PEFT
• Quantisation: 4-bit NF4 (QLoRA) during fine-tuning

---

Limitations

• Corpus is weighted toward formal, urban Sinhala (news, Wikipedia, government sources)
• Dialectal, informal, and rural Sinhala registers are underrepresented
• Some shards have inconsistent metadata fields due to multi-source scraping

---

Licence

This dataset is released under Creative Commons Attribution 4.0 (CC BY 4.0). You are free to use, share, and adapt this data for any purpose with attribution.

---

Citation

@misc{amarasinghe2026serendib,
  title   = {Serendib LLM: A Sinhala Large Language Model for NLP Benchmarks and Poem Generation},
  author  = {Amarasinghe, Chamaka},
  year    = {2026},
  school  = {University of Central Lancashire},
  url     = {https://huggingface.co/Chamaka8}
}

---

Contact

Chamaka Amarasinghe University of Central Lancashire — CO3008 Honours Degree Project 2025–2026 GitHub: chamakarochana LinkedIn: chamaka-amarasinghe-b54904211