Heron-NVILA-Lite-15B

Heron-NVILA-Lite-15B is a vision language model trained for Japanese, based on the NVILA-Lite architecture.

Model Overview

• Developer: Turing Inc.
• Vision Encoder: paligemma-siglip-so400m-patch14-448
• Projector: mlp_downsample_3x3_fix
• LLM: Qwen2.5-14B-Instruct
• Supported Languages: Japanese, English

Setup

# I have confirmed that 4.46.0 and 4.49.0 also work. Other versions of Transformer may also work, but I have not tested them.
pip install transformers==4.45.0 accelerate opencv-python torchvision einops pillow
pip install git+https://github.com/bfshi/scaling_on_scales.git

Usage

from transformers import AutoConfig, AutoModel

model_path = "turing-motors/Heron-NVILA-Lite-15B"

# you can use config
config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
model = AutoModel.from_config(config, trust_remote_code=True, device_map="auto")

# or directly from_pretrained
model = AutoModel.from_pretrained(model_path, trust_remote_code=True, device_map="auto")

# show chat_template
print(model.tokenizer.chat_template)

# examples generate with raw text
response = model.generate_content(["こんにちは"])
print(response)
print("---" * 40)

# examples generate with text + image
from PIL import Image
import requests
url = "http://images.cocodataset.org/val2017/000000039769.jpg"
image = Image.open(requests.get(url, stream=True).raw).convert("RGB")
response = model.generate_content([image, "画像を説明してください。"])
print(response)
print("---" * 40)

# examples generate using generation_config
from PIL import Image
import requests
from transformers import GenerationConfig
generation_config = {
    "max_new_tokens": 512,
    "temperature": 0.5,
    "do_sample": True,
}
generation_config = GenerationConfig(**generation_config)
url = "http://images.cocodataset.org/val2017/000000039769.jpg"
image = Image.open(requests.get(url, stream=True).raw).convert("RGB")
response = model.generate_content(
    [image, "画像を説明してください。"],
    generation_config=generation_config
)
print(response)
print("---" * 40)

# examples generate with text + image + text + image + text
from PIL import Image
import requests
url_list = [
    "https://images.unsplash.com/photo-1694831404826-3400c48c188d?q=80&w=2070&auto=format&fit=crop&ixlib=rb-4.0.3&ixid=M3wxMjA3fDB8MHxwaG90by1wYWdlfHx8fGVufDB8fHx8fA%3D%3D",
    "https://images.unsplash.com/photo-1693240876439-473af88b4ed7?q=80&w=1974&auto=format&fit=crop&ixlib=rb-4.0.3&ixid=M3wxMjA3fDB8MHxwaG90by1wYWdlfHx8fGVufDB8fHx8fA%3D%3D"
]
images = [
   Image.open(requests.get(url, stream=True).raw).convert("RGB") for url in url_list
]
response = model.generate_content([
    images[0],
    "これは日本の画像です",
    images[1],
    "これはオーストリアの画像です",
    "各画像の違いを説明して"])
print(response)
print("---" * 40)

Training Summary

Stage	Training	Data Sources	Samples
Stage1	Projector	Japanese image text pairs, LLaVA-Pretrain	1.1M
Stage2	Projector, LLM	Filtered MOMIJI (CC-MAIN-2024-46, CC-MAIN-2024-51, CC-MAIN-2025-05)	13M
		Japanese image text pairs (subset), Japanese interleaved data (subset), mmc4-core (subset), coyo-700m (subset), wikipedia_ja, llava_pretrain_ja, stair_captions	20M
Stage3	Vision Encoder, Projector, LLM	llava-instruct-v1_5-en-subset-358k, llava-instruct-ja, japanese-photos-conv, ja-vg-vqa, synthdog-ja (subset), ai2d, synthdog-en, sherlock	1.1M

Evaluation

I used llm-jp-eval-mm for this evaluation. Scores for models other than Heron-NVILA-Lite and Sarashina2-Vision-14B were taken from llm-jp-eval-mm leaderboard as of March 2025 and the Asagi website. Heron-NVILA-Lite and Sarashina2-Vision-14B were evaluated using llm-as-a-judge with "gpt-4o-2024-05-13". Sarashina2-Vision-14B was evaluated on the official blog using "gpt-4o-2024-08-06"; please note that due to differing evaluation conditions, the results for Sarashina2-Vision-14B should be treated as reference only.

Model	LLM Size	Heron-Bench overall LLM (%)	JA-VLM-Bench-In-the-Wild LLM (/5.0)	JA-VG-VQA-500 LLM (/5.0)
Heron-NVILA-Lite-1B	0.5B	45.9	2.92	3.16
Heron-NVILA-Lite-2B	1.5B	52.8	3.52	3.50
Heron-NVILA-Lite-15B	14B	59.6	4.2	3.82
LLaVA-CALM2-SigLIP	7B	43.3	3.15	3.21
Llama-3-EvoVLM-JP-v2	8B	39.3	2.92	2.96
VILA-jp	13B	57.2	3.69	3.62
Asagi-14B	13B	55.8	3.44	3.84
Sarashina2-Vision-14B	13B	50.9	4.1	3.43
Qwen2-VL 7B Instruct	7B	55.5	3.61	3.6
GPT-4o	-	87.6	3.85	3.58

Risks and Limitations

This model is experimental and has not been thoroughly calibrated for ethical compliance or legal standards. Caution is advised for sensitive applications.

License

• Model weights are licensed under Apache License 2.0.
• Users must comply with OpenAI terms of use due to the inclusion of GPT-4-generated synthetic data.

Acknowledgements

This model is based on results obtained from a project, JPNP20017, subsidized by the New Energy and Industrial Technology Development Organization (NEDO).

I would like to acknowledge the use of the following open-source repositories:

• VILA
• llm-jp-eval-mm