OpenThai-SystemOne

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.

Quickstart

pip install "git+https://github.com/iapp-technology/openthai-systemone"      # or: pip install openthai-systemone

from 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 8000

Model card with benchmark tables: https://huggingface.co/iapp/OpenThai-SystemOne. Design rationale: docs/PLAN.md.

Layout

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

Pipeline

Dev

uv venv -p 3.12 .venv && source .venv/bin/activate
uv pip install -e ".[server,train,dev]"
pytest -q

Synthetic data needs an OpenAI-compatible endpoint:

export SYNTH_BASE_URL=http://localhost:11436/v1 SYNTH_MODEL=openthai2.0-qwen3.8-27b SYNTH_API_KEY=...
python scripts/03b_synth_generate.py --out data/synth --n 600000 --concurrency 32