monoboard/thai-land-tax-full-triplets

Thai Land & Buildings Tax — Full Triplet Dataset

Dataset Name: monoboard/thai-land-tax-full-triplets Language: Thai (th) Tasks: Legal Retrieval • RAG • Contrastive Learning • Triplet Loss • Embedding Training

Overview

This dataset provides a legally verified retrieval corpus for training Thai-language retrieval models under the Land and Buildings Tax Act (B.E. 2562) and related regulations. Each example includes:

• A legal question (query)
• One or more oracle passages (pos)
• One or more hard-negative distractors (neg)

This enables:

• Multi-positive dense retrieval
• Hard-negative contrastive learning
• Triplet training
• Legal citation alignment
• RAG grounding for Thai law

Files Included

File	Description
thai_land_tax_full_triplets.jsonl	Complete dataset (multi-pos/multi-neg)

Canonical format: multi-pos/multi-neg JSONL

Data Format

Canonical Format (recommended)

{
  "query": "ที่ดินที่กำลังก่อสร้างถือเป็นประเภทใด?",
  "pos": [
    "ประกาศกระทรวงการคลังและมหาดไทย ข้อ 3 ...",
    "พระราชบัญญัติภาษีที่ดินและสิ่งปลูกสร้าง พ.ศ. 2562 มาตรา 5 ..."
  ],
  "neg": [
    "มาตรา 37(4) ที่ดินรกร้าง",
    "พระราชกฤษฎีกา 2564 อัตราเพื่อการพาณิชย์ ..."
  ]
}

Loading the Dataset

Multi-pos/multi-neg (recommended)

from datasets import load_dataset

ds = load_dataset("monoboard/thai-land-tax-full-triplets", split="train")
print(ds[0])

Expand into Triplets

from sentence_transformers import InputExample
from datasets import load_dataset

ds = load_dataset("monoboard/thai-land-tax-full-triplets", split="train")

def expand(example):
    triplets = []
    for pos in example["pos"]:
        for neg in example["neg"]:
            triplets.append(InputExample(texts=[example["query"], pos, neg]))
    return triplets

triplet_list = []
for ex in ds:
    triplet_list.extend(expand(ex))

Using with SentenceTransformers (Triplet Loss)

from sentence_transformers import SentenceTransformer, losses
from torch.utils.data import DataLoader

model = SentenceTransformer("BAAI/bge-m3")

train_dataloader = DataLoader(triplet_list, batch_size=8, shuffle=True)
loss = losses.TripletLoss(model)

model.fit(
    train_objectives=[(train_dataloader, loss)],
    epochs=3,
    warmup_steps=500,
)

Using for RAG Retrieval (BGE-M3 Example)

from datasets import load_dataset
from sentence_transformers import SentenceTransformer

ds = load_dataset("monoboard/thai-land-tax-full-triplets", split="train")

model = SentenceTransformer("BAAI/bge-m3")

queries = [x["query"] for x in ds]
candidates = list({c for item in ds for c in item["pos"] + item["neg"]})

embeddings = model.encode(candidates, batch_size=32)