humanizerth-thai-12B-GGUF

GGUF Quantizations of humanizerth-thai-12B for llama.cpp, Ollama, LM Studio, and Cross-Platform Inference

humanizerth-thai-12B-GGUF provides official GGUF quantizations of humanizerth, a specialized model designed to remove artificial, robotic AI phrasing from Thai text and restore natural, human-written conversational cadence while preserving 100% of facts, names, figures, and prices.

Fine-tuned on top of jialinyyzz/humanizer and google/gemma-4-12B with 10,000 authentic pre-LLM human Thai text pairs (from Wongnai reviews, Pantip discussions, and Wisesight social media).

Also available for Apple Silicon via Apple MLX:

chantra/humanizerth-thai-12B-Q8_0-MLX (~11.8 GB)
chantra/humanizerth-thai-12B-Q4_0-MLX (~6.3 GB)

📦 Provided GGUF Files

File Name	Quant Method	File Size	Recommended RAM	Description
`humanizerth-thai-12B-Q8_0.gguf`	`Q8_0`	~12.2 GB	16–24 GB	Maximum fidelity, near-lossless Thai nuances
`humanizerth-thai-12B-Q4_K_M.gguf`	`Q4_K_M`	~7.3 GB	10–14 GB	Recommended default - balanced speed and quality
`humanizerth-thai-12B-Q4_0.gguf`	`Q4_0`	~6.8 GB	8–12 GB	Fastest inference, lowest memory footprint

📊 Benchmark & Evaluation Results (100 Held-Out Samples)

Metric	Base Humanizer (Zero-Shot)	humanizerth (Q8_0 GGUF)	humanizerth (Q4_K_M GGUF)	Status
Ref Token Similarity (Human Jaccard)	48.50%	94.36%	94.10%	🟢 100% Win Rate
Median Verbatim AI Reuse (Lower is better)	0.0820	0.0350	0.0550	🟢 Passed (< 0.22)
Hard Fact Recall (Numbers/Prices/Dates)	100.0%	100.0% (100/100)	100.0% (100/100)	🟢 Zero Fact Loss
Language Purity (CJK Leakage)	0 / 100	0 / 100 (0%)	0 / 100 (0%)	🟢 Clean Thai Output

🚀 How to Use

1. Using `llama-cli` (llama.cpp)

llama-cli -m humanizerth-thai-12B-Q4_K_M.gguf -ngl 99 -c 4096 -n 512 -p "Rewrite the text below so it reads like a person wrote it, not a language model.

Reorganize it as you see fit. Vary sentence length on purpose. Cut hedging,
throat-clearing, and any sentence that only announces what comes next.
Prefer the concrete word over the abstract one. It is fine to sound uneven.

Every fact, number, unit, date, name and quotation must survive unchanged.

จากการประเมินคุณภาพของร้านคาเฟ่เปิดใหม่ย่านอารีย์ พบว่าเมนูมัทฉะลาเต้ ราคา 160 บาท มีรสชาติกลมกล่อม...

### Rewritten:

"

2. Using with Ollama

Create a file named Modelfile:

FROM ./humanizerth-thai-12B-Q4_K_M.gguf

TEMPLATE """Rewrite the text below so it reads like a person wrote it, not a language model.

Reorganize it as you see fit. Vary sentence length on purpose. Cut hedging,
throat-clearing, and any sentence that only announces what comes next.
Prefer the concrete word over the abstract one. It is fine to sound uneven.

Every fact, number, unit, date, name and quotation must survive unchanged.

{{ .Prompt }}

### Rewritten:

"""

PARAMETER stop "###"
PARAMETER stop "<end_of_turn>"
PARAMETER temperature 0.3

Build and run:

ollama create humanizerth -f Modelfile
ollama run humanizerth "ข้อความภาษา AI ที่ต้องการแปลง..."

3. Using with LM Studio / LocalAI

Search for chantra/humanizerth-thai-12B-GGUF or import the downloaded .gguf file.
Load the model with GPU acceleration enabled.
Use the system instruction and completion format: `### Rewritten:

🎨 Persona Steering via Prefix Priming

Because the model is completion-based, you can steer tone into Gen Z / Social Media Review style by appending a prefix right after `### Rewritten:

Food/Cafe Review: `### Rewritten:

ทุกคนนน คือปังมากกกก `

Travel Review: `### Rewritten:

ทริปนี้คือปังไม่ไหวววว ค่าาา `

Tech Review: `### Rewritten:

แกรรร บอกเลยว่าโคตรปังงง `

📜 Credits & Citation

@misc{humanizerth2026,
  author = {Chantra / Thanawit},
  title = {humanizerth-thai-12B-GGUF: Faithful Thai Text Humanization in GGUF},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/chantra/humanizerth-thai-12B-GGUF}}
}

chantra/humanizerth-thai-12B-GGUF