Kanitakorn-Qwen3-8B-SFT-v5 (NLU specialist)
A LoRA adapter on top of Qwen/Qwen3-8B, fifth iteration. Beats Pathumma-1.0-7B on 3 of 4 NLU benchmarks (Pathumma's published suite).
Headline: Pathumma-style NLU (apples-to-apples, our pipeline, n=30/benchmark)
| benchmark | Kanitakorn-v5 | Kanitakorn-v1 | Pathumma-1.0-7B (pub) | OpenThaiGPT-1.5-7B (pub) | result |
|---|---|---|---|---|---|
| xcopa_th | 0.867 | 0.833 | 0.83 | 0.854 | WIN +3.7pp |
| belebele_th | 0.833 | 0.767 | 0.7777 | 0.7944 | WIN +5.5pp |
| xnli_th | 0.633 | 0.20 | 0.4011 | 0.397 | WIN +23.2pp 🔥 |
| wisesight | 0.10 | 0.30 | 0.4129 | 0.5024 | regresses |
Other benchmarks (our pipeline, n=15)
| benchmark | v5 (no /no_think) |
|---|---|
| ThaiExam (n=33 spot-check) | 0.424 |
| OpenThaiEval MCQ | 0.333 |
| IFEval-TH strict | 0.733 |
| MT-Bench-TH | 5.20 |
| MATH500-TH | 0.40 |
| HotpotQA EM | 0.200 |
When to use v5 vs v1
- v5 (this model): when NLU tasks matter — xcopa, belebele, xnli, classification, multi-task reasoning
- v1: when ThaiExam-style academic MCQ is the priority (0.503 vs v5's 0.424)
- For chat-heavy use, v4 is best (MT-Bench 5.87)
Training recipe
- Base: Qwen/Qwen3-8B
- LoRA r=64, alpha=128, all-linear, NEFTune α=5, use_rslora=True
- Unsloth 2026.5.2 (2.4× speedup)
- lr=3e-5 cosine, 1 epoch, max_steps=2400
- bs=8 × grad_accum=2 = effective batch 16
- 37,715 records, 45% Thai-native:
- 4,215 Kanitakorn original
- 10,000 WangchanThaiInstruct
- 3,000 wisesight (sentiment)
- 4,000 Tulu-3 wildchat / OASST / no_robots
- 3,000 Capybara
- 8,500 numinamath + personahub_math + gsm8k
- 3,000 evol_codealpaca + personahub_code
- 2,000 Typhoon-S-instruct supplements
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base = "Qwen/Qwen3-8B"
adapter = "Jnx03/kanitakorn-qwen3-8b-sft-v5"
tok = AutoTokenizer.from_pretrained(base, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(model, adapter).merge_and_unload()
# IMPORTANT: do NOT inject /no_think for this SFT model.
# /no_think is for the raw Qwen3-8B base only.
# Let v5 reason naturally on Thai academic questions.
Honest caveats
Qwen3-8B raw is already very strong on Thai NLU — our re-measurement shows raw beats Pathumma on belebele (0.933 vs 0.7777). v5's wins on xcopa/belebele are partly explained by the strong base; the xnli +23pp win IS the SFT's contribution.
wisesight regression (-20pp from v1): v5's training has Wildchat + Capybara English chat that may bias toward "positive" sentiment outputs. R6 (more wisesight + Thaweewat QA) addresses this.
ThaiExam regression (-7.9pp from v1): diversity-vs-specialization trade-off. v1 is more concentrated on Thai academic; v5 sacrifices some academic for broader NLU.
See https://huggingface.co/datasets/Jnx03/kanitakorn-th-sft for full development log and methodology.