Downloads · 30 days
0
Jnx03/kanitakorn-qwen3-8b-sft-v1
kanitakorn-qwen3-8b-sft-v1 is a text generation model from Jnx03. Use it when you need the model to write or continue text. The card lists the license as odc-by.
A LoRA adapter on top of Qwen/Qwen3-8B, trained on a 4,147-record Thai-focused SFT corpus spanning 8 benchmark families (AIME-TH, MATH500-TH, LiveCodeBench-TH, OpenThaiEval, MT-Bench-TH, IFEval-TH, HotpotQA-TH, plus a…
Downloads · 30 days
0
Access
Public
Updated May 18, 2026
Repo size
710 MB
Likes
0
Public
Click a slice to open those files.
.safetensors698 MB · 98%
From the Hugging Face model README
A LoRA adapter on top of Qwen/Qwen3-8B, trained on a 4,147-record Thai-focused SFT corpus
spanning 8 benchmark families (AIME-TH, MATH500-TH, LiveCodeBench-TH, OpenThaiEval, MT-Bench-TH,
IFEval-TH, HotpotQA-TH, plus a teacher_loop_th family).
On the same evaluation pipeline (greedy decoding, /no_think for Qwen3 thinking-mode, n=15-30/family, LLM judge = google/gemini-2.5-flash-lite):
| benchmark | Kanitakorn-8B-SFT | Typhoon-S-8B | delta |
|---|---|---|---|
| ThaiExam (n=165) | 0.503 | 0.467 | +3.6pp |
| MATH500-TH | 0.533 | 0.400 | +13.3pp |
| HotpotQA EM | 0.200 | 0.133 | +6.7pp |
| AIME24 | 0.033 | 0.000 | +3.3pp |
| OpenThaiEval MCQ | 0.267 | 0.267 | 0 |
| AIME25 | 0.000 | 0.000 | 0 |
| LiveCodeBench | 0.000 | 0.000 | 0 |
| IFEval-TH strict | 0.667 | 0.733 | -6.7pp |
| MT-Bench-TH | 4.50 | 7.80 | -3.30 |
Qwen/Qwen3-8Bfrom transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base = "Qwen/Qwen3-8B"
adapter = "Jnx03/kanitakorn-th-sft" # this repo (LoRA only)
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)
model = model.merge_and_unload() # optional: merge for faster inference
messages = [
{"role": "user", "content": "วันนี้คุณรู้สึกอย่างไร?"}
]
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tok(text, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=300, do_sample=False, pad_token_id=tok.eos_token_id)
print(tok.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
Qwen/Qwen3-8B (MT-Bench 6.53) instead.@misc{kanitakorn2026,
title = {Kanitakorn-Qwen3-8B-SFT: A Thai benchmark-targeted SFT on Qwen3-8B},
author = {Jnx03},
year = {2026},
url = {https://huggingface.co/datasets/Jnx03/kanitakorn-th-sft},
}