A LoRA adapter on top of Qwen/Qwen3-4B-Instruct-2507, trained on the same Thai-focused 4,147-record corpus as the 8B variant. Designed for deployment to mobile / edge / Typhoon-S-like small-device targets.

Results (LLM judge enabled, n=15/family)

benchmark	Kanitakorn-4B-SFT	Kanitakorn-8B-SFT (for ref)	Typhoon-S-8B (for ref)
ThaiExam (n=165)	0.418	0.503	0.467
MATH500-TH	0.467	0.533	0.400
IFEval-TH strict	0.733	0.667	0.733
IFEval-TH loose	0.767	0.744	0.789
OpenThaiEval MCQ	0.400	0.267	0.267
AIME24	0.067	0.033	0.000
AIME25	0.067	0.000	0.000
HotpotQA EM	0.200	0.200	0.133
MT-Bench-TH	3.6	4.5	7.8

Kanitakorn-4B (half the params) beats Typhoon-S-8B on:

AIME24 +6.7pp
AIME25 +6.7pp
MATH500-TH +6.7pp
HotpotQA +6.7pp
OpenThaiEval MCQ +13.3pp
IFEval-TH strict TIE

Training recipe (identical to 8B)

LoRA r=64, α=128, dropout=0.05, target_modules='all-linear', NEFTune α=5
lr=1e-4 cosine, 2 epochs, bf16, gradient_checkpointing
1× A100, ~3h wall time

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "Qwen/Qwen3-4B-Instruct-2507"
adapter = "Jnx03/kanitakorn-qwen3-4b-sft-v1"
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()