Open-source Thai + English System One decision model: a Qwen3.5-0.8B text tower whose LM head is replaced by a
256-way slot softmax, continued-pretrained on Thai, and trained to answer typed questions (choice / score /
noul) about an arbitrary state in a single forward pass with calibrated probabilities. The request/response contract
mirrors TypeSafe AI's POST /v1/systemone (Jev), so SDK code written for TypeSafe can point at this model.
pip install "git+https://github.com/iapp-technology/openthai-systemone" # or: pip install openthai-systemonefrom openthai_systemone import SystemOneClient, Choice, Score, Noul
client = SystemOneClient("iapp/OpenThai-SystemOne") # CUDA, MPS or CPU
resp = client.system_one(
state={"ticket": "โดนหักเงินซ้ำสองครั้ง ขอเงินคืนด่วน โทรไปสามรอบแล้วไม่มีใครรับ"},
questions={
"department": Choice(instructions="ทีมใดควรรับผิดชอบ", criteria={"billing": "การเงิน/คืนเงิน", "technical": "ระบบใช้งานไม่ได้", "sales": None}),
"frustration": Score(instructions="ลูกค้าหงุดหงิดแค่ไหน", criteria=["ใจเย็น", "หงุดหงิดแต่สุภาพ", "โกรธมาก"]),
"refund": Noul(instructions="ลูกค้าขอเงินคืนอย่างชัดเจนหรือไม่"),
},
)
print(resp.answers["department"].choice, resp.answers["department"].probabilities)
print(resp.answers["frustration"].score, resp.answers["refund"].noul)HTTP server with the same JSON contract as TypeSafe's POST /v1/systemone:
OPENTHAI_SYSTEMONE_MODEL=iapp/OpenThai-SystemOne uvicorn openthai_systemone.server:app --port 8000Model card with benchmark tables: https://huggingface.co/iapp/OpenThai-SystemOne.
openthai_systemone/ pip package: types (Choice/Score/Noul), formatting (prompt + slots), modeling (SlotHead), client, server
scripts/00..07_*.py one script per pipeline stage (see below)
configs/*.yaml CPT / SFT / calibration hyper-parameters (1x H100)
tests/ CPU unit tests
| stage | script | where | output |
|---|---|---|---|
| 0 surgery | scripts/00_prepare_base.py |
CPU | base/qwen3.5-0.8b-text (vision encoder dropped) |
| 1 CPT data | scripts/01_cpt_data.py |
CPU, needs internet | data/cpt/shard-*.bin (5B Thai-heavy tokens) |
| 1 CPT | scripts/02_cpt_train.py --config configs/cpt.yaml |
1x H100, ~1 day | runs/cpt/latest |
| 2 decision data | scripts/03_decision_data.py, scripts/03b_synth_generate.py |
CPU + LLM endpoint | data/decision/*.jsonl, data/synth/*.jsonl |
| 2 SFT | scripts/04_decision_train.py --config configs/sft.yaml |
1x H100, ~0.5 day | runs/sft/latest |
| 3 calibration | scripts/04_decision_train.py --config configs/calib.yaml --init runs/sft/latest |
1x H100, hours | runs/calib/latest |
| 4 eval | scripts/06_eval.py --model runs/calib/latest |
GPU | runs/eval.json |
| 5 release | scripts/07_export_release.py --ckpt runs/calib/latest --out release/... --push iapp/OpenThai-SystemOne |
– | HF Hub |
uv venv -p 3.12 .venv && source .venv/bin/activate
uv pip install -e ".[server,train,dev]"
pytest -qSynthetic data needs an OpenAI-compatible endpoint:
export SYNTH_BASE_URL=http://localhost:8000/v1 SYNTH_MODEL=<any-openai-compatible-chat-model> SYNTH_API_KEY=<key-or-none>
python scripts/03b_synth_generate.py --out data/synth --n 600000 --concurrency 32
Training, evaluation and the free hosted API for this model run on NVIDIA H100 GPUs generously provided by Siam AI Corporation. Thank you for backing open Thai AI.