Modelpublic

iapp/OpenThai-SystemOne-MLX-4bit

sourceHugging Faceapache-2.0updated 11d agoView on Hugging Face

0likes23downloads

Model Card

OpenThai-SystemOne — mlx-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 tower including the token embeddings (MLX quantizes the embedding table too). The 256-slot decision head and the per-type temperatures stay in fp32 (head.safetensors). Quantization therefore only perturbs the hidden state the head reads.

Format: MLX 4-bit affine (group 64) for Apple Silicon (mlx-lm). mlx-lm runs the tower; the included client applies the decision head on the final hidden states. Size: 424 MB.

Measured on a MacBook Pro M3 Max: 4-bit ≈ 19 ms per 3-question Thai decision (the PyTorch model on MPS: ~150 ms).

Usage

bash

pip install mlx-lm torch transformers safetensors pydantic
huggingface-cli download iapp/OpenThai-SystemOne-MLX-4bit --local-dir openthai-mlx

python

import sys; sys.path.insert(0, "openthai-mlx")
from openthai_systemone.mlx_client import MLXSystemOneClient
c = MLXSystemOneClient("openthai-mlx")
r = c.system_one("ร้านนี้อาหารอร่อยมาก แต่รอนานเกือบชั่วโมง", {"sentiment": {"type": "choice", "instructions": "ความรู้สึก",
                 "criteria": {"บวก": None, "ลบ": None, "กลาง": None}}})
print(r.answers["sentiment"].choice, r.answers["sentiment"].probabilities)

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

Macro: public 72.9 (original 74.3), Thai 79.0 (original 80.1).

subset	bf16 original	this	Δ
public 13-subset bench
aegis2 (noul)	83.2	82.0	-1.2
boolq (noul)	79.7	77.7	-2.0
civil_comments (noul)	79.0	80.3	+1.3
helpsteer2 (score)	41.6	41.6	+0.0
massive-de-DE (choice)	88.3	79.4	-8.9
massive-en-US (choice)	88.3	79.7	-8.6
multinli (choice)	89.0	88.0	-1.0
paws (noul)	94.0	92.4	-1.6
pubmedqa (choice)	64.0	64.0	+0.0
squad2 (noul)	89.3	87.0	-2.3
summeval-consistency (score)	75.0	80.6	+5.6
summeval-relevance (score)	21.7	25.0	+3.3
vitaminc-dev (choice)	72.5	70.6	-1.8
macro, public 13-subset bench	74.3	72.9	-1.3
Thai held-out / eval sets
banking77 (choice)	59.1	52.1	-7.0
contrastive_th (choice)	80.7	80.4	-0.3
contrastive_th (noul)	83.5	83.9	+0.4
contrastive_th (score)	78.6	78.6	+0.0
massive_th (choice)	90.6	87.5	-3.1
prachathai (choice)	98.3	98.8	+0.5
prachathai (noul)	93.4	92.7	-0.8
sib200_th (choice)	77.9	77.0	-1.0
wisesight (choice)	48.9	46.5	-2.4
wongnai (score)	64.5	65.2	+0.8
xlam_tools (choice)	99.4	99.4	+0.0
xnli_th (choice)	79.8	79.0	-0.8
xnli_th (noul)	86.8	86.0	-0.8
macro, Thai held-out / eval sets	80.1	79.0	-1.1

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.