benchmark	v7	R6	v1	v5
ThaiExam	0.479	0.467	0.503	0.424
IFEval-TH strict	0.800	0.800	0.667	0.733
MT-Bench-TH	6.067	6.80	4.50	5.20
MATH500-TH	0.267	0.333	0.533	0.40
OpenThaiEval MCQ	0.333	0.333	0.267	0.333
HotpotQA EM	0.067	0.067	0.200	0.200

Pathumma extended NLU (n=30)

benchmark	v7	R6	Pathumma-pub
xcopa_th	—	0.833	0.83
belebele_th	—	0.867	0.7777
xnli_th	—	0.567	0.4011
wisesight	—	0.70	0.4129

Training recipe

python

DPOConfig(
    loss_type="apo_zero",   # Anchored preference optimization, reference-free
    beta=2.5,
    learning_rate=1e-6,
    num_train_epochs=2,
    per_device_train_batch_size=2, gradient_accumulation_steps=16,
    max_length=2048, max_prompt_length=1800,
    lr_scheduler_type="cosine", warmup_ratio=0.1,
)

Base: R6 LoRA adapter (continuing training)
Preference data: 469 codex-distilled Thai (chosen, rejected) pairs
30 steps total (~6 min on A100-80GB)
Reward margin reached 30-38 (model learning preferences)

Pref data composition (469 pairs)

chat_quality (60): high-quality vs low-quality Thai chat
instruct_follow (60): constraint compliance
multi_turn (60): dialogue continuations
reasoning (60): correct CoT vs flawed
politeness (60): correct vs wrong register
factual (60): accurate vs subtly wrong facts
refusal (60): helpful vs over-cautious
conciseness (49): concise vs bloated

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-8B", torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(model, "Jnx03/kanitakorn-qwen3-8b-sft-v7-simpo").merge_and_unload()
# Same as R6 — do NOT inject /no_think for SFT models.