BowLingual-12B-v1

BowLingual-12B-v1

BowLingual-12B-v1 is a fine-tuned version of google/gemma-4-12B-it, intended primarily for natural English-to-Thai conversational translation.

It targets informal messages such as LINE and Messenger chats, where literal textbook translation often sounds stiff or unnatural.

Intended use

• English-to-Thai conversational translation
• Informal, natural-sounding Thai messages
• Preserving tone, directness, and emotional intensity
• Prompt-driven translation pipelines

Usage

Install the dependencies:

pip install -U transformers torch accelerate

Example:

from transformers import AutoModelForMultimodalLM, AutoProcessor

model_id = "stenml/BowLingual-12B-v1"

processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForMultimodalLM.from_pretrained(
    model_id,
    dtype="auto",
    device_map="auto",
)

messages = [
    {
        "role": "user",
        "content": "Translate into natural casual Thai: I'm almost home.",
    }
]

inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
    add_generation_prompt=True,
    enable_thinking=False,
).to(model.device)

output = model.generate(**inputs, max_new_tokens=256)
generated = output[0][inputs["input_ids"].shape[-1]:]
print(processor.decode(generated, skip_special_tokens=True))

Default generation settings

• Temperature: 1.0
• Top-p: 0.95
• Top-k: 64

Translation applications may benefit from lower-temperature or deterministic decoding, depending on the prompt and desired amount of stylistic variation.

Files

• model.safetensors: full model weights
• tokenizer.json: Gemma 4 tokenizer
• tokenizer_config.json: tokenizer and response parsing metadata
• chat_template.jinja: canonical Gemma 4 chat template
• processor_config.json: multimodal processor configuration

The Q4_K_M GGUF release is available at stenml/BowLingual-12B-v1-GGUF.

Evaluation

No formal benchmark results are currently published for this release. Test it against representative conversational messages before production use.

Limitations

• Primarily intended for English-to-Thai conversational translation
• May produce inaccurate, unnatural, or overly literal translations
• May hallucinate information absent from the source
• Not validated for legal, medical, financial, or safety-critical translation
• Multimodal behavior has not been formally evaluated after fine-tuning
• Training-data and hyperparameter details are not currently published

License

Apache License 2.0, following the Gemma 4 base model metadata.