iapp/OpenThai-SystemOne-Ollama

OpenThai-SystemOne v0.3 for Ollama

OpenThai-SystemOne is an open Thai + English System One decision model (0.8B, Apache-2.0). It does not generate text: given a state (text or JSON) and typed questions it returns probabilities: choice between named options, noul (yes/no) and score on an ordered scale.

This repo is its Ollama build for Ollama's System One API (POST /v1/systemone, Ollama ≥ 0.35), the same API Ollama serves Nimble and Tev1 with. Everything runs on your machine; no API key.

ollama pull iapp/openthai-systemone                       # ollama.com: 0.8b (= 0.8b-q8_0), 0.8b-q4_K_M, 0.8b-bf16
ollama pull hf.co/iapp/OpenThai-SystemOne-Ollama:Q8_0     # the same files from this repo

The curl example below uses the hf.co name; with the ollama.com pull, use "model": "iapp/openthai-systemone".

curl http://localhost:11434/v1/systemone -d '{
  "model": "hf.co/iapp/OpenThai-SystemOne-Ollama:Q8_0",
  "state": {"ticket": "ลูกค้าแจ้งว่าโดนหักเงินซ้ำสองครั้ง ขอเงินคืนด่วน โทรมาสามรอบแล้ว"},
  "questions": {
    "department": {"type": "choice", "instructions": "ทีมใดควรรับผิดชอบ",
                   "criteria": {"billing": "การเงิน/ค่าบริการ", "technical": "ระบบใช้งานไม่ได้", "sales": null}},
    "frustration": {"type": "score", "instructions": "ลูกค้าหงุดหงิดแค่ไหน",
                    "criteria": ["ใจเย็น", "หงุดหงิดแต่สุภาพ", "โกรธมาก"]},
    "refund_requested": {"type": "noul", "instructions": "ลูกค้าขอเงินคืนอย่างชัดเจนหรือไม่"}
  }
}'

{"model": "hf.co/iapp/OpenThai-SystemOne-Ollama:Q8_0",
 "answers": {
   "department": {"type": "choice", "choice": "billing", "probabilities": {"billing": 0.9515, "technical": 0.0334, "sales": 0.0151}, "confidence": 0.7959},
   "frustration": {"type": "score", "score": 1.8771, "legend": {"0": "ใจเย็น", "1": "หงุดหงิดแต่สุภาพ", "2": "โกรธมาก"},
                   "probabilities": {"0": 0.0216, "1": 0.0797, "2": 0.8987}, "confidence": 0.6537},
   "refund_requested": {"type": "noul", "noul": 0.9805}},
 "usage": {"input_tokens": 741, "output_tokens": 4}}

(Probabilities rounded. The whole request is part of every question's prompt, so the same question can score slightly differently next to other questions.)

How this build differs from the main repo

Ollama does not run OpenThai-SystemOne's 256-slot decision head. For each question it renders one chat prompt (the whole request as JSON plus Requested field: "<name>", with the model's Qwen3.5 chat template and thinking off) and reads the next-token probabilities of the answer letters A–Z. The weights here are therefore v0.3 fine-tuned for that prompt:

• 3,000 steps (192k questions) on the v0.3 training mix, rendered exactly as Ollama 0.35 renders them (a byte-exact port of Ollama's decision/systemone.go, checked against Go) and tokenized by llama.cpp, as Ollama's runner does.
• Loss = cross-entropy over the question's candidate letters only, which is the softmax Ollama computes.
• One temperature, fitted on held-out records, is folded into the final norm (Ollama always scores at temperature 1).
• The GGUF files are a plain Qwen3.5 (qwen35) text model with tied embeddings; the Modelfile / system file sets the system prompt the model was trained with and num_ctx 8192.

Compared with the main repo's own API (pip install openthai-systemone): Ollama allows 2–26 options per question (the main API: 255), runs one prompt per question (the shared prefix is cached), and has no order-invariant mode and no abstain answer.

Evaluation (through Ollama 0.35)

All columns were run by us through Ollama's /v1/systemone on the same records: the first 800 of each set, keeping only records whose questions have ≤ 26 options (Ollama's limit; drops banking77 and the 60-way MASSIVE-th intents). The first column is the original v0.3 weights with their 256-slot head on the same records, for reference. choice / noul = accuracy, score = exact level. Harness: scripts/25_competitor_eval.py --model ollama:<name> --max-options 26 in the GitHub repo.

Public 13 subsets (Bespoke Nimble's public benchmark)

set (type, n)	v0.3, main repo's API	this repo, Q8_0	this repo, Q4_K_M	Tev1 0.8B	Tev1 4B	Nimble 9B
aegis2 (noul, n=250)	83.2	82.0	81.2	60.8	80.8	83.2
boolq (noul, n=300)	79.7	79.7	81.0	77.7	85.3	86.3
civil_comments (noul, n=300)	79.0	76.0	77.3	76.7	73.3	78.0
helpsteer2 (score, n=250)	41.6	39.2	40.0	33.2 (1 err)	36.8 (1 err)	33.6
massive-de-DE (choice, n=350)	88.6	88.0	87.4	69.7	83.1	83.1
massive-en-US (choice, n=350)	89.1	88.9	88.6	78.9	85.4	84.0
multinli (choice, n=299)	88.6	86.0	84.3	75.6	92.0	90.0
paws (noul, n=250)	94.0	92.8	92.8	65.2	82.4	74.0
pubmedqa (choice, n=250)	64.0	65.6	65.6	61.2	74.4	77.2
squad2 (noul, n=299)	89.3	89.0	86.3	70.9	76.3	74.2
summeval-consistency (score, n=144)	75.0	76.4	74.3	84.0	79.9	81.9
summeval-relevance (score, n=240)	21.7	28.7	27.9	13.8	50.0	48.3
vitaminc-dev (choice, n=599)	72.5	74.1	75.0	68.8	74.3	79.0
macro	74.3	74.3	74.0	64.3	74.9	74.8

Thai sets (8 sets, 12 rows; wisesight and SIB-200 held out of training)

set (type, n)	v0.3, main repo's API	this repo, Q8_0	this repo, Q4_K_M	Tev1 0.8B	Tev1 4B	Nimble 9B
contrastive_th (choice, n=296)	80.7	83.8	82.4	75.7	92.2	93.2
contrastive_th (noul, n=248)	83.5	84.7	84.3	79.0	94.8	97.2
contrastive_th (score, n=56)	78.6	78.6	75.0	60.7	91.1	83.9
massive_th (choice, n=588)	94.6	92.7	92.3	71.6	88.8	90.8
prachathai (choice, n=413)	98.5	97.8	97.6	56.7	61.7	61.5
prachathai (noul, n=1568)	93.4	95.2	95.0	68.8	74.2	66.3
sib200_th (choice, n=204)	77.9	78.9	76.5	83.3	86.3	88.7
wisesight (choice, n=800)	49.0	49.9	49.8	40.8	48.0	48.5
wongnai (score, n=800)	64.5	64.0	62.3	38.5	56.1	52.1
xlam_tools (choice, n=800)	99.4	99.4	99.4	90.1	97.1	97.2
xnli_th (choice, n=800)	79.8	79.2	78.5	66.9	76.5	76.0
xnli_th (noul, n=800)	86.8	85.9	85.6	20.9	38.4	84.0
macro	82.2	82.5	81.6	62.7	75.4	78.3

Latency (median end to end through Ollama, one model loaded, idle H100, Thai requests): 1 question 23 ms (Q8_0), 23 ms (Q4_K_M), 26 ms (BF16); 3 questions 136 / 127 / 142 ms. Same setup: Tev1 0.8B 22 / 102 ms, Tev1 4B 66 / 406 ms, Nimble 9B 69 / 392 ms. Ollama runs one prompt per question, reusing the shared prefix.

BF16 scores the same as Q8_0 (public 74.3, Thai 82.5).

Files

file	size	public / Thai macro through Ollama
OpenThai-SystemOne-v0.3-Ollama-Q8_0.gguf	812 MB	74.3 / 82.5 (recommended)
OpenThai-SystemOne-v0.3-Ollama-Q4_K_M.gguf	529 MB	74.0 / 81.6
OpenThai-SystemOne-v0.3-Ollama-BF16.gguf	1517 MB	74.3 / 82.5
Modelfile, Modelfile.Q4_K_M, Modelfile.BF16	–	for ollama create from a local file
system, params	–	system prompt and num_ctx 8192, applied by ollama pull hf.co/...

Modelfile builds the same model from a local file (ollama create openthai-systemone -f Modelfile); system and params are what ollama pull hf.co/... applies.

Limits

• Up to 26 options per question (Ollama). For more options, bucket them, or use the main repo's API (255 options).
• Not a chat model: ollama run will produce text, but the model was only trained to answer /v1/systemone prompts.
• A small model: use confidence and route low-confidence decisions to a bigger model or a person.

License

Apache-2.0. Built by iApp Technology / OpenThai on Qwen3.5-0.8B-Base (Apache-2.0). Not affiliated with TypeSafe AI, Bespoke Labs or Together AI.