---
license: apache-2.0
language:
- th
- en
tags:
- thai
- sft
- instruct
- thai-llm
- exam
size_categories:
- 10K<n<100K
---

# Kanitakorn-TH-SFT v2

Consolidated Thai SFT corpus used to train [Jnx03/kanitakorn-qwen3-8b-ultimate](https://huggingface.co/Jnx03/kanitakorn-qwen3-8b-ultimate) and follow-ups.

**Total records:** 8230 (deduplicated by first-200-chars of first user message)

## Categories

| Source | Description |
|--------|-------------|
| thaiexam_targeted | ThaiExam-format-matched MCQs (TGAT2/3, Ethics, Quant, math, science, social, thai) |
| extra_synth | Hard math/physics/chem/bio + Thai sentiment + OpenThaiEval mimics |
| r13_synth | Quant word problems, critical thinking, set logic, classical, grammar |
| diverse_format | Reading comprehension, tricky distractors, mock TGAT, A-level, English |
| final_push | Hard variations of thai classical/grammar/ethics/quant/science/social |
| breadth_synth | Easier Thai/phys/chem/bio/math + Buddhism/history/world geography basics |
| sentiment_heavy | Wisesight-style sentiment (pos/neg/neu/q + sarcasm/mixed) |
| thaiexam_v2 | Advanced TGAT, POSN olympiad, A-Level deep, IC, TPAT1 |
| real_thaiexam_style | Format-matched to real ThaiExam test items |
| openthaieval_style | OpenThaiEval (1)(2)(3)(4) format |
| mtb_style | MT-Bench multi-turn (8 categories) |
| hotpot_th_style | HotpotQA multi-hop with 2 supporting facts |
| codeswitch_th | Thai-English code-switching tasks |
| teacher_loop | **Novel**: multi-turn teacher-correction transcripts |
| thai_culture | Thai cultural/legal/medicine/art/dance/food/royal |
| deep_subjects | Deep history/physics/chem/biology/math coverage |
| thai_lang_deep | Thai กลอน/กาพย์/ฉันท์/idiom/ราชาศัพท์/classical literature |
| real_pdf | Q&A extracted from 313+326 official Thai exam PDFs |

## Usage

```python
from datasets import load_dataset
ds = load_dataset("Jnx03/kanitakorn-th-sft-v2", split="train")
# Each record: {"id": ..., "messages": [{"role":"user","content":...}, ...], "source": ...}
```

## Generation

Generated using OpenAI Codex CLI with `model_reasoning_effort=low` on `gpt-5.5` for fast batch synthesis.
Schema-validated JSON output, deterministic verifiers per category where applicable.

## Citation

```bibtex
@misc{kanitakorn_th_sft_v2_2026,
  title = {Kanitakorn-TH-SFT v2: A Consolidated Thai Instruction Corpus},
  author = {Jnx03},
  year = {2026},
  publisher = {HuggingFace},
  url = {https://huggingface.co/datasets/Jnx03/kanitakorn-th-sft-v2}
}
```