README.md

---
language:
- si
license: mit
size_categories:
- 100K<n<1M
task_categories:
- text-generation
pretty_name: UltraChat-Sinhala
source_datasets:
- HuggingFaceH4/ultrachat_200k
tags:
- sinhala
- machine-translation
- nllb
- instruction-tuning
- sft
configs:
- config_name: default
  data_files:
  - split: train_sft
    path: data/train_sft.parquet
  - split: test_sft
    path: data/test_sft.parquet
  - split: train_gen
    path: data/train_gen.parquet
  - split: test_gen
    path: data/test_gen.parquet
dataset_info:
  features:
  - name: prompt
    dtype: string
  - name: prompt_id
    dtype: string
  - name: messages
    list:
    - name: content
      dtype: string
    - name: role
      dtype: string
  splits:
  - name: train_sft
    num_bytes: 3220638735
    num_examples: 207831
  - name: test_sft
    num_bytes: 356835963
    num_examples: 23106
  - name: train_gen
    num_bytes: 3112868195
    num_examples: 255974
  - name: test_gen
    num_bytes: 342615462
    num_examples: 28300
  download_size: 1568756226
  dataset_size: 7032958355
---

# Dataset Card for UltraChat-Sinhala

## Dataset Description

UltraChat-Sinhala is a **Sinhala (සිංහල) machine translation of
[HuggingFaceH4/ultrachat_200k](https://huggingface.co/datasets/HuggingFaceH4/ultrachat_200k)**,
built to supervised-fine-tune Sinhala large language models. It preserves the
original dataset's structure, splits, and `prompt_id`s, so it is a drop-in
Sinhala counterpart to the English source.

The English dialogues were translated with
[NLLB-200-3.3B](https://huggingface.co/facebook/nllb-200-3.3B)
(`eng_Latn → sin_Sinh`) and then put through a Sinhala-specific cleaning
pipeline (conjunct/ZWJ repair, masking-leak repair, filtering). The dataset
contains **515,211 dialogues / ≈528M Sinhala tokens** (SinLLaMA tokenizer).

## Dataset Creation

1. **Translation.** Each dialogue turn was translated en→si with NLLB-200-3.3B
   (greedy decoding). To respect the model's 512-token limit, turns were
   sentence-segmented with the line layout preserved on reassembly. Spans that
   must not be translated — fenced/inline code, URLs, e-mail addresses, HTML
   tags, LaTeX/maths and markdown links — were masked before translation and
   restored afterwards, so they pass through verbatim.

2. **Sinhala conjunct (ZWJ) repair.** NLLB's SentencePiece normaliser strips the
   Zero-Width Joiner (`U+200D`) from Sinhala conjunct clusters, emitting a space
   instead (e.g. `ප් ර` for `ප්‍ර`). A lexicon-gated restorer — built from a
   Sinhala corpus plus the tokenizer's vocabulary — re-inserts the joiner only
   where attested. After repair, **~99.6–99.9% of dialogues carry conjunct
   joiners**.

3. **Masking-leak repair.** The placeholder used to mask the protected spans was
   itself corrupted by SentencePiece (its rare brackets were stripped),
   displacing each masked span to the end of its message with a stray digit left
   behind. These were re-inserted in their correct positions by re-aligning each
   message to its English source (≈83% reconstructed exactly in place; the rest
   re-appended without any content loss).

4. **Filtering.** Dialogues containing empty/whitespace-only turns were dropped
   (a mid-conversation turn cannot be removed without breaking the
   user/assistant alternation); `prompt_id`s were de-duplicated; and one
   `prompt_id` shared between the SFT and GEN sets was removed for global
   id-uniqueness.

5. **Train/test split.** Membership is taken **verbatim from the original
   ultrachat_200k split** by `prompt_id`, so the Sinhala split is identical to
   the English source (≈10% test). It is **leak-checked**: no `prompt_id` and no
   identical dialogue appears in both train and test, in either SFT or GEN.

## Dataset Structure

Like the source, the dataset has four splits, suitable for:

* Supervised fine-tuning (`sft`).
* Generation ranking (`gen`) via techniques like rejection sampling or PPO.

| split | examples | tokens |
|:----------|---------:|------------------:|
| train_sft | 207,831  | 248,133,641 |
| test_sft  | 23,106   | 27,467,232  |
| train_gen | 255,974  | 227,432,690 |
| test_gen  | 28,300   | 25,035,938  |
| **total** | **515,211** | **528,069,501** |

Tokens are raw message-content tokens (the SinLLaMA tokenizer, vocab 139,336, has
no chat template; a real SFT run adds a small per-turn special-token overhead).

The dataset is stored in parquet (zstd), schema-identical to the source:

```
{
    "prompt": "ආහාර පිසීමේ ව්‍යාපාරයක් සඳහා වට්ටෝරු පොතක් නිර්මාණය කරන්න. ...",
    "prompt_id": "7d86ffeefdea030c92138e0b964c304508bfebed4b23261c8a741630823e6f96",
    "messages": [
        {
            "role": "user",
            "content": "ආහාර පිසීමේ ව්‍යාපාරයක් සඳහා වට්ටෝරු පොතක් නිර්මාණය කරන්න. ..."
        },
        {
            "role": "assistant",
            "content": "නම: රසවත් ආහාර පිසීමේ රහස්: රසවත් හා ලස්සන ආහාර පිසීමේ වට්ටෝරු පොත ..."
        },
        {
            "role": "user",
            "content": "ඔයා මේ වෙනකම් හදපු වට්ටෝරු පොත නම් නියමයි. ඔයාට පුලුවන්ද තව විස්තර එකතු කරන්න ..."
        },
        {
            "role": "assistant",
            "content": "..."
        }
    ]
}
```

## Quality and Limitations

Full-scan checks on the released data: **100% valid JSON**, correct
user/assistant alternation, no empty turns, no duplicate `prompt_id`s.
~**0.017%** of turns remain in Latin script (overwhelmingly fenced code, which is
preserved by design), and ~**0.5%** of longer turns show NLLB repetition
artifacts.

This is **machine translation without human post-editing**. Expect:

- translationese and occasional mistranslation, especially on idioms, named
  entities, and technical content;
- code, identifiers, URLs and maths intentionally left in their original form;
- a minority of turns with NLLB repetition loops (~0.5%);
- residual orthographic edge cases not covered by the conjunct lexicon.

It is intended for instruction-tuning and research, not as a gold-standard
reference translation.

## Licensing

Released under the **MIT License**, following the source dataset
`HuggingFaceH4/ultrachat_200k`. The Sinhala text was produced by machine
translation with NLLB-200.