Downloads · 30 days
22
27% of all-time downloads
xCloudinfo/Llama-xReadable-11B-zhTW-OCR-v1
Llama-xReadable-11B-zhTW-OCR-v1 is a image-text-to-text model from xCloudinfo. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. The card lists the license as llama3.2.
Downloads · 30 days
22
27% of all-time downloads
All-time downloads
81
Public
Parameters
10.7B
21.4 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors21.3 GB · 100%
From the Hugging Face model README
Built with Llama
云碩科技 · xCloudinfo · 系列:Meta Llama 3.2 Vision · 云碩繁體中文OCR
以 meta-llama/Llama-3.2-11B-Vision-Instruct 為基底,由云碩科技在自有AI算力資源池上微調的繁體中文(台灣)OCR模型。專注台灣場景的文字辨識——公文、票據、機房與售前筆記等中英混寫內容,可地端部署、資料不外流。xReadable:讓文件被讀懂。
定位:印刷/掃描文件辨識已達可用水準;手寫辨識為實驗性預覽(experimental preview),持續迭代中。
評測皆以真實拍攝/掃描影像計算字元錯誤率(CER,越低越好),非合成樣本。
| 場景 | CER | 說明 |
|---|---|---|
| 印刷/掃描繁中文件(真實照片120張) | 0.93% | 可用水準 |
| 手寫繁中筆記(真實照片,短篇業務筆記) | 64.1%(4張) | experimental preview,尚未達生產可用 |
| 手寫繁中筆記(長篇多句內容) | 142.2%(16張) | experimental preview,長文本仍會出現內容幻覺 |
本repo目前僅提供 bf16 safetensors(transformers原生格式),尚未提供GGUF量化版本。
mllama 架構,llama.cpp官方目前不支援轉換或載入(上游issue自2024-09未解決)。這代表 llama.cpp / llama-server / LM Studio / GPT4All等以llama.cpp為底層的工具目前都無法運行本模型(市面上少數自稱支援的mllama GGUF是用Ollama私有fork轉出,僅能在該環境下運作,與標準llama.cpp生態不相容)。上游一旦支援,本repo會補上官方相容的GGUF版本。transformers(Python)與 vLLM 皆原生支援 mllama 架構,可直接載入本模型。bf16權重約 21GB,建議:
device_map="auto" 搭配多卡分攤,或自行做int8/int4量化(bitsandbytes)降低需求。from transformers import MllamaForConditionalGeneration, AutoProcessor
from PIL import Image
import torch
model_id = "xCloudinfo/Llama-xReadable-11B-zhTW-OCR-v1"
model = MllamaForConditionalGeneration.from_pretrained(
model_id, torch_dtype=torch.bfloat16, device_map="auto"
).eval()
processor = AutoProcessor.from_pretrained(model_id)
image = Image.open("your_document.jpg").convert("RGB")
prompt = "請辨識圖片中的所有文字,逐字完整輸出,只輸出文字本身。"
messages = [{"role": "user", "content": [{"type": "image"}, {"type": "text", "text": prompt}]}]
text = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
inputs = processor(text=[text], images=[image], return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=768, do_sample=False)
print(processor.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
vllm serve xCloudinfo/Llama-xReadable-11B-zhTW-OCR-v1 \
--dtype bfloat16 \
--max-model-len 4096
啟動後以OpenAI相容API(圖片以 image_url 傳入 messages)呼叫即可。
本模型為Meta Llama 3.2的衍生模型,散布須遵守 Llama 3.2 Community License Agreement。以下依該授權條款聲明:
meta-llama/Llama-3.2-11B-Vision-Instruct(Meta Platforms, Inc.)。Llama 3.2為Meta之財產,Llama Materials 之權利與所有權屬Meta。Llama-xReadable-11B-zhTW-OCR-v1)。LICENSE 檔);使用、重製、散布本模型或其衍生物,須提供該授權副本並保留本聲明。Built with Llama. 由 云碩科技xCloudinfo於自有AI算力資源池微調、散布;基底模型之權利屬Meta Platforms, Inc.。
云碩科技(xCloudinfo Corp. Limited)專注於地端私有化人工智慧解決方案。xVITA是云碩自研模型的品牌名稱。
其他模型請見 https://hf.co/xCloudinfo。
Built with Llama
xCloudinfo (云碩科技) · Series: Meta Llama 3.2 Vision · xCloudinfo Traditional Chinese OCR
A Traditional Chinese (Taiwan) OCR model built on meta-llama/Llama-3.2-11B-Vision-Instruct and fine-tuned by xCloudinfo on its own AI compute pool. It focuses on text recognition for Taiwan scenarios — the mixed Chinese/English content of official documents, receipts, server-room and pre-sales notes, and the like — and can be deployed on-premises with no data leaving the site. xReadable: making documents readable.
Positioning: printed/scanned document recognition has reached a usable level; handwriting recognition is an experimental preview, still being iterated.
All evaluations compute the character error rate (CER, lower is better) on real photographed/scanned images, not synthetic samples.
| Scenario | CER | Notes |
|---|---|---|
| Printed/scanned Traditional Chinese documents (120 real photos) | 0.93% | Usable level |
| Handwritten Traditional Chinese notes (real photos, short business notes) | 64.1% (4 photos) | Experimental preview, not yet production-usable |
| Handwritten Traditional Chinese notes (long, multi-sentence content) | 142.2% (16 photos) | Experimental preview, long text still produces content hallucinations |
This repo currently provides only bf16 safetensors (the native transformers format) and does not yet provide a GGUF quantized version.
mllama architecture used by Llama 3.2 Vision is not currently supported for conversion or loading by llama.cpp (the upstream issue has been open since 2024-09). This means tools based on llama.cpp such as llama.cpp / llama-server / LM Studio / GPT4All cannot currently run this model (the few mllama GGUFs on the market that claim support were converted with a private Ollama fork, work only in that environment, and are incompatible with the standard llama.cpp ecosystem). Once upstream adds support, this repo will add an officially compatible GGUF version.transformers (Python) and vLLM both natively support the mllama architecture and can load this model directly.The bf16 weights are about 21GB; recommendations:
device_map="auto" to split across multiple cards, or quantize to int8/int4 yourself (bitsandbytes) to lower the requirement.from transformers import MllamaForConditionalGeneration, AutoProcessor
from PIL import Image
import torch
model_id = "xCloudinfo/Llama-xReadable-11B-zhTW-OCR-v1"
model = MllamaForConditionalGeneration.from_pretrained(
model_id, torch_dtype=torch.bfloat16, device_map="auto"
).eval()
processor = AutoProcessor.from_pretrained(model_id)
image = Image.open("your_document.jpg").convert("RGB")
prompt = "請辨識圖片中的所有文字,逐字完整輸出,只輸出文字本身。"
messages = [{"role": "user", "content": [{"type": "image"}, {"type": "text", "text": prompt}]}]
text = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
inputs = processor(text=[text], images=[image], return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=768, do_sample=False)
print(processor.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
vllm serve xCloudinfo/Llama-xReadable-11B-zhTW-OCR-v1 \
--dtype bfloat16 \
--max-model-len 4096
After startup, call it via the OpenAI-compatible API (pass images into messages as image_url).
This model is a derivative of Meta Llama 3.2, and distribution is subject to the Llama 3.2 Community License Agreement. The following statement is made per that license:
meta-llama/Llama-3.2-11B-Vision-Instruct (Meta Platforms, Inc.). Llama 3.2 is the property of Meta, and the rights and ownership of the Llama Materials belong to Meta.Llama-xReadable-11B-zhTW-OCR-v1).LICENSE file in this repo); using, reproducing, or distributing this model or its derivatives requires providing the license copy and retaining this statement.Built with Llama. Fine-tuned and distributed by xCloudinfo (云碩科技) on its own AI compute pool; rights in the base model belong to Meta Platforms, Inc.
xCloudinfo Corp. Limited (云碩科技) builds on-premises, private AI solutions. xVITA is xCloudinfo's in-house model brand.
More models: https://hf.co/xCloudinfo