---
license: cc-by-4.0
pretty_name: Bintulu LLM Dataset (text-to-text, vector-ready)
language:
  - en
  - ms
tags:
  - text-to-text
  - text2text
  - bintulu
  - sarawak
  - borneo
  - tourism
  - energy-hub
  - bilingual
  - instruction-tuning
  - retrieval
task_categories:
  - text-generation
  - translation
language_details: Bahasa Malaysia (ms) + English (en); `mix` rows are task prompts in either language
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train.jsonl
---

# Bintulu LLM Dataset

A text-to-text (prompt → answer) corpus about Bintulu, Sarawak — its national
parks, beaches, energy industry, history, food, transport and guesthouse
hospitality. Built for two uses:

1. **Fine-tuning** a text-to-text LLM (instruction, QA, dialogue, translation,
   NLI, MCQ).
2. **Embedding/retrieval** — every row's `text` field (canonically equal to
   `input_text`) is the embeddable surface for a vector store.

## Schema

Each JSONL row:

| field         | type   | meaning                                             |
|---------------|--------|-----------------------------------------------------|
| `input_text`  | string | prompt / source sentence (embed with `text`)        |
| `target_text` | string | expected answer / completion                        |
| `text`        | string | canonical embeddable surface (`== input_text`)      |
| `meta`        | dict   | `id` (`bintulu-t2t-#####`), `category`, `lang`, `source` |

## Counts

- **Rows: 1681** unique `(input_text, target_text)` pairs (seeded, deduplicated).
- Categories: translation (753), itinerary (300), sarawak_malay (282), knowledge_qa (69), multiple_choice (55), dialogue (49), recommendation (44), definition (39), instruction (35), nli (29), fact_check (26).
- Languages: en 646, mix 502, **ms 533 (≥500 — includes 282 rows of Bahasa Melayu Sarawak dialect under the `sarawak_malay_*` categories)**.
- Deterministic: regenerate identically with
  `python build_bintulu.py` (seed `20260905`).

## Load locally

```bash
pip install datasets
from datasets import load_dataset
ds = load_dataset("ctaxnagomi/bintulu-llm-dataset", split="train")
print(ds[0])
```

or directly:

```bash
python -c "import json; rows=[json.loads(l) for l in open('data/train.jsonl')]; print(len(rows))"
```

## Embedding submission (vector store ingestion)

Requirements for submitting this corpus to an embedding model:

- **Model**: multilingual recommended for `en`/`ms` rows, e.g.
  `BAAI/bge-m3` (1024-dim) or `jinaai/jina-embeddings-v2-base-en`
  (T5-style, 8192-token window) via the Hugging Face Inference API.
- **Embedded surface**: the `text` field, but for retrieval tasks embed the
  *query* and *document* with the same prompt style (`text` == `input_text`).
- **Normalisation**: L2-normalise vectors before storing; retrieval uses cosine
  similarity.
- **Truncation**: truncate to the model window (512 tokens for BGE-family /
  Sentence-BERT, 8192 for Jina v2). Truncate from the *left* for long prompts.
- **Batching**: batch ≤ 64 rows, keep the HTTP connection alive, and retry with
  exponential backoff on HTTP 429 / 503.
- **Precision**: store `float32`; row order must match `train.jsonl`.
- **Auth**: send `Authorization: Bearer {{YourHFToken}}` (replace the placeholder
  with your HF read token; never commit the token).

See [`scripts/embed.py`](scripts/embed.py) for a ready-to-run ingestion script.

## Grounding sources

Wikipedia (Bintulu, Similajau National Park), Sarawak Tourism Board, Sarawak
Forestry Corporation, Bintulu Development Authority (BDA), and sector-verified
facts collected for the Kim Guesthouse project. Sarawak Malay dialect rows are
grounded in Asmah Haji Omar (2015), Wikipedia "Sarawak Malay" (phonology and
vocabulary, Nabilah Bolhassan 2019), the Borneo Dictionary Melayu Sarawak
wordlists, and the DangoSarawak community glossary. Generated by a
deterministic seeded builder (`ctecx-ai-grounded`, `ctecx-ai-grounded-sarawak`).