change	v1	v2	hypothesis
Dataset size	4,147 records	23,715 records (4,147 + 19,500 filtered)	does broader corpus help?
Augmentation	none	autoif Thai-IF 8k + math 8.5k + code 3k from `typhoon-ai/typhoon-s-instruct-post-training`	targets IFEval + MATH gaps
Recipe	LoRA r=64 alpha=128 all-linear, NEFTune α=5	same	isolates corpus effect
Training	2 epochs ~1,652 steps	1 epoch, 1,800 max steps	one pass over the larger corpus
Loss masking	text-mode loss-on-all (TRL default)	same (text-mode, the loss-on-prompt bug remains)	v3 tests the recipe fix

Results

WARNING — incomplete evaluation. Eval pipeline was too slow (~0.04 rec/s on the v2-SFT model). Killed at record 28/552 to free GPU for v3 then v4. We retain partial scores; the full eval was not completed.

Partial v2 scores (n=28, mostly one family):

benchmark	partial score	sample
(single-family partial — not meaningful as a comparison)	—	n=28

For v1 head-to-head numbers (n=15-312), see https://huggingface.co/Jnx03/kanitakorn-qwen3-8b-sft-v1

Training recipe

python

LoraConfig(
    r=64, lora_alpha=128, target_modules='all-linear',
    lora_dropout=0.05, bias='none', task_type='CAUSAL_LM',
)
SFTConfig(
    num_train_epochs=1, max_steps=1800,
    per_device_train_batch_size=2, gradient_accumulation_steps=8,
    learning_rate=1e-4, bf16=True,
    warmup_ratio=0.05, weight_decay=0.01,
    lr_scheduler_type='cosine',
    neftune_noise_alpha=5.0,
    gradient_checkpointing=True,
)

Data composition (verified, no test contamination)

source	records	tag
Original Kanitakorn (gap-weighted Thai benchmark families)	4,215	thai-targeted
ai2-adapt-dev/autoif (filtered from typhoon-s-instruct)	8,000	thai-IF
numinamath_tir + personahub_math + tulu_open_math + tulu_personas_math	8,500	math CoT
personahub_code_v2 + evol_codealpaca_heval_decontaminated	3,000	code

License: ODC-By (inherited from typhoon-ai/typhoon-s-instruct-post-training).

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "Qwen/Qwen3-8B"
adapter = "Jnx03/kanitakorn-qwen3-8b-sft-v2"
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()

Honest disclosure

v2 was the broader-corpus experiment. We did NOT complete benchmark evaluation before pivoting to v3 (research-driven recipe fix) and v4 (Unsloth-accelerated + Typhoon-2-inspired). The v2 checkpoint is published for transparency. v1 remains the recommended baseline release; once v3/v4 are fully benchmarked one of them may supersede v1.

See https://huggingface.co/datasets/Jnx03/kanitakorn-th-sft REPORT_HONEST.md and RESEARCH_LOG.md for the full development log.