Modelpublic

iapp/OpenThai-SystemOne-bnb-4bit

sourceHugging Faceapache-2.0updated 15d agoView on Hugging Face

0likes15downloads

Model Card

OpenThai-SystemOne — bnb-4bit

OpenThai-SystemOne is an open Thai + English System One decision model: one forward pass answers typed questions (choice over up to 255 options, ordinal score, yes/no noul) about a text / JSON state with calibrated probabilities, no text generation. It is a Qwen3.5-0.8B text tower (Thai continued pre-training) plus a 256-slot decision head. This repo is a quantization of v0.3 (commit f3709948).

What is quantized: the Linear layers of the tower. The token embeddings, the 256-slot decision head and the per-type temperatures stay in bf16. Quantization therefore only perturbs the hidden state the head reads.

Format: bitsandbytes NF4 4-bit (double quant). bitsandbytes checkpoint, stays quantized in GPU memory (CUDA).

Size: 767 MB (bf16 original: 1,509 MB).

Usage

bash

pip install torch transformers safetensors pydantic && pip install bitsandbytes

python

from transformers import AutoModel, AutoTokenizer        # trust_remote_code files are in this repo
model = AutoModel.from_pretrained("iapp/OpenThai-SystemOne-bnb-4bit", trust_remote_code=True)
# or with the pip package (git+https://github.com/iapp-technology/openthai-systemone):
from openthai_systemone import SystemOneClient
c = SystemOneClient("iapp/OpenThai-SystemOne-bnb-4bit")
r = c.system_one("ร้านนี้อาหารอร่อยมาก แต่รอนานเกือบชั่วโมง", {"sentiment": {"type": "choice", "instructions": "ความรู้สึก",
                 "criteria": {"บวก": None, "ลบ": None, "กลาง": None}}})

Accuracy vs the bf16 original (same records, single option order, first 800 per set)

Macro: public 73.6 (original 74.3), Thai 79.7 (original 80.1).

subset	bf16 original	this	Δ
public 13-subset bench
aegis2 (noul)	83.2	82.8	-0.4
boolq (noul)	79.7	78.7	-1.0
civil_comments (noul)	79.0	81.0	+2.0
helpsteer2 (score)	41.6	42.8	+1.2
massive-de-DE (choice)	88.3	85.7	-2.6
massive-en-US (choice)	88.3	86.3	-2.0
multinli (choice)	89.0	87.0	-2.0
paws (noul)	94.0	92.8	-1.2
pubmedqa (choice)	64.0	62.8	-1.2
squad2 (noul)	89.3	87.0	-2.3
summeval-consistency (score)	75.0	73.6	-1.4
summeval-relevance (score)	21.7	26.2	+4.6
vitaminc-dev (choice)	72.5	69.8	-2.7
macro, public 13-subset bench	74.3	73.6	-0.7
Thai held-out / eval sets
banking77 (choice)	59.1	58.2	-0.9
contrastive_th (choice)	80.7	79.7	-1.0
contrastive_th (noul)	83.5	81.9	-1.6
contrastive_th (score)	78.6	82.1	+3.6
massive_th (choice)	90.6	90.5	-0.1
prachathai (choice)	98.3	98.8	+0.5
prachathai (noul)	93.4	93.4	+0.0
sib200_th (choice)	77.9	75.0	-2.9
wisesight (choice)	48.9	49.5	+0.6
wongnai (score)	64.5	63.9	-0.6
xlam_tools (choice)	99.4	99.4	+0.0
xnli_th (choice)	79.8	77.9	-1.9
xnli_th (noul)	86.8	86.2	-0.5
macro, Thai held-out / eval sets	80.1	79.7	-0.4

Notes

—Scores are single-option-order accuracy on the first 800 records of each set (scripts/06_eval.py --limit 800), the same records for the original and the quantization. score subsets report exact level accuracy.
—Base model, data, training and the full benchmark tables: iapp/OpenThai-SystemOne.
—License Apache-2.0 (same as the base). Built by iApp Technology / OpenThaiGPT.