Downloads · 30 days
0
Jnx03/kanitakorn-qwen3-8b-sft-v6
kanitakorn-qwen3-8b-sft-v6 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 Qwen/Qwen3-8B, sixth iteration. Biggest corpus (49,650 records) including 2,538 codex-distilled (GPT-5.5) Thai synthetic records — the most ambitious recipe.
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 Qwen/Qwen3-8B, sixth iteration. Biggest corpus (49,650 records) including
2,538 codex-distilled (GPT-5.5) Thai synthetic records — the most ambitious recipe.
| benchmark | v6 | v1 (Thai academic) | v5 (NLU) | Typhoon-S-8B (paper) |
|---|---|---|---|---|
| ThaiExam | 0.467 | 0.503 | 0.424 | NF |
| IFEval-TH strict | 0.800 | 0.667 | 0.733 | 0.7645 |
| MT-Bench-TH | 6.800 | 4.50 | 5.20 | 7.89 |
| MATH500-TH | 0.333 | 0.533 | 0.40 | 0.6593 |
| OpenThaiEval MCQ | 0.333 | 0.267 | 0.333 | 0.671 |
| AIME24 | 0.000 | 0.033 | 0 | NF |
| AIME25 | 0.000 | 0 | 0 | NF |
| HotpotQA EM | 0.067 | 0.200 | 0.20 | 0.370 |
| LiveCodeBench | 0.000 | 0 | 0 | NF |
| benchmark | v6 | v5 | Pathumma (pub) | OTG (pub) |
|---|---|---|---|---|
| xcopa_th | — | 0.867 | 0.83 | 0.854 |
| belebele_th | — | 0.833 | 0.7777 | 0.7944 |
| xnli_th | — | 0.633 | 0.4011 | 0.397 |
| wisesight | — | 0.10 | 0.4129 | 0.5024 |
Qwen/Qwen3-8Bairesearch/WangchanThaiInstructThaweewat/instruct-qa-thai-combined (TyDi/XQuAD/MKQA NLU)airesearch/wangchanx-seed-free-synthetic-instruct-thai-120k (MCQ-heavy)pythainlp/han-instruct-dataset-v4.0pythainlp/wisesight_sentiment (sentiment classification)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-v6").merge_and_unload()
# IMPORTANT: do NOT inject /no_think for SFT models — it breaks Thai academic eval.
See https://huggingface.co/datasets/Jnx03/kanitakorn-th-sft for full methodology.