Downloads · 30 days
478
25% of all-time downloads
xCloudinfo/xVITA-Embed-300M-zhTW-GGUF
xVITA-Embed-300M-zhTW-GGUF is a sentence similarity model from xCloudinfo. Use it when you need a score for how close two texts are. The card lists the license as gemma.
xVITA-Embed-300M-zhTW(云碩科技自製繁體中文檢索embedding,DRCD test hit@1 86.0、超越bge-m3 / multilingual-e5-large)的GGUF量化版,適用llama.cpp與ollama。
Downloads · 30 days
478
25% of all-time downloads
All-time downloads
1.9K
Public
Repo size
3.8 GB
Likes
0
Public
Click a slice to open those files.
.gguf1.9 GB · 100%
From the Hugging Face model README
xVITA-Embed-300M-zhTW(云碩科技自製繁體中文檢索embedding,DRCD test hit@1 86.0、超越bge-m3 / multilingual-e5-large)的GGUF量化版,適用llama.cpp與ollama。
完整精度權重(safetensors)為私有發行,如有微調或研究合作需求請聯繫云碩科技。
2026-07-18重要修復:本日以前的所有GGUF檔因轉檔工具與sentence-transformers 5.6模組命名不相容, 缺少模型的Dense投影層,向量空間與safetensors版不一致(cosine相似度失效)。請重新下載本日後的檔案。 現版本已逐一與safetensors版驗證向量一致(f16 cosine = 1.0、Q8_0 = 0.9998),並已把mean pooling 寫入GGUF metadata,llama.cpp無須再手動指定
--pooling。
| 版本 | context | 檔案 | 大小 | 適用 |
|---|---|---|---|---|
| 標準版(v2) | 8192 | xcloud-emb-zhtw-f16.gguf | 584 MB | 一般RAG、段落級檢索(短文分數最佳86.0) |
| xcloud-emb-zhtw-Q8_0.gguf | 313 MB | |||
| 長文特化版 | 8192 | xcloud-emb-zhtw-8k-f16.gguf | 584 MB | 長文件整篇檢索、法規/報告全文(長文分數最佳84.7) |
| xcloud-emb-zhtw-8k-Q8_0.gguf | 313 MB |
短段落檢索(3,493查詢 / 1,000段落庫):
| 模型 | hit@1 | hit@5 | hit@10 |
|---|---|---|---|
| 標準版v2(0.3B) | 86.0 | 97.4 | 98.7 |
| multilingual-e5-large(0.56B) | 82.8 | 95.3 | 97.0 |
| BAAI/bge-m3(0.57B) | 82.2 | 95.6 | 97.2 |
長文整篇檢索(378篇完整文章 / 3,493查詢):
| 模型 | hit@1 | hit@5 | hit@10 |
|---|---|---|---|
| 長文特化版 | 84.7 | 94.5 | 96.6 |
| 標準版v2 | 79.2 | 91.2 | 94.3 |
| BAAI/bge-m3(8k原生) | 77.5 | 90.3 | 93.4 |
長文特化版在完整文章檢索上勝過bge-m3原生8k達7.2個百分點。
# 標準版(段落級 RAG)
llama-server -m xcloud-emb-zhtw-Q8_0.gguf --embeddings --port 8080
# 長文特化版(整篇長文件檢索)
llama-server -m xcloud-emb-zhtw-8k-Q8_0.gguf --embeddings --port 8080
查詢端與文檔端請使用EmbeddingGemma規格prompt:
task: search result | query: {你的問題}title: none | text: {文檔內容}完整評測、訓練細節與誠實聲明見主repo model card。
底模依 Gemma Terms of Use 授權,本量化模型之使用同受其約束。
Gemma合規聲明
Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms
註:GGUF檔名維持
xcloud-emb-*前綴不變,以免既有下載腳本與ollama設定失效;模型本身即xVITA-Embed-300M-zhTW。
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。
The GGUF quantization of xVITA-Embed-300M-zhTW (an xCloudinfo in-house Traditional Chinese retrieval embedding, DRCD test hit@1 86.0, surpassing bge-m3 / multilingual-e5-large), for llama.cpp and ollama.
The full-precision weights (safetensors) are a private release; for fine-tuning or research-collaboration needs, please contact xCloudinfo.
2026-07-18 important fix: all GGUF files from before this date were, due to a naming incompatibility between the conversion tool and sentence-transformers 5.6 modules, missing the model's Dense projection layer, so their vector space did not match the safetensors version (cosine similarity was broken). Please re-download the files dated after this day. The current version has been verified vector-by-vector against the safetensors version (f16 cosine = 1.0, Q8_0 = 0.9998), and mean pooling has been written into the GGUF metadata, so llama.cpp no longer needs
--poolingspecified manually.
| Version | context | File | Size | Suited to |
|---|---|---|---|---|
| Standard (v2) | 8192 | xcloud-emb-zhtw-f16.gguf | 584 MB | general RAG, passage-level retrieval (best short-text score 86.0) |
| xcloud-emb-zhtw-Q8_0.gguf | 313 MB | |||
| Long-document specialized | 8192 | xcloud-emb-zhtw-8k-f16.gguf | 584 MB | whole-document retrieval of long documents, full text of regulations/reports (best long-text score 84.7) |
| xcloud-emb-zhtw-8k-Q8_0.gguf | 313 MB |
Short-passage retrieval (3,493 queries / 1,000-passage pool):
| Model | hit@1 | hit@5 | hit@10 |
|---|---|---|---|
| Standard v2 (0.3B) | 86.0 | 97.4 | 98.7 |
| multilingual-e5-large (0.56B) | 82.8 | 95.3 | 97.0 |
| BAAI/bge-m3 (0.57B) | 82.2 | 95.6 | 97.2 |
Whole-article long-document retrieval (378 complete articles / 3,493 queries):
| Model | hit@1 | hit@5 | hit@10 |
|---|---|---|---|
| Long-document specialized | 84.7 | 94.5 | 96.6 |
| Standard v2 | 79.2 | 91.2 | 94.3 |
| BAAI/bge-m3 (native 8k) | 77.5 | 90.3 | 93.4 |
On complete-article retrieval the long-document specialized version beats bge-m3's native 8k by 7.2 percentage points.
# 標準版(段落級 RAG)
llama-server -m xcloud-emb-zhtw-Q8_0.gguf --embeddings --port 8080
# 長文特化版(整篇長文件檢索)
llama-server -m xcloud-emb-zhtw-8k-Q8_0.gguf --embeddings --port 8080
On both the query side and the document side, use the EmbeddingGemma-spec prompt:
task: search result | query: {你的問題}title: none | text: {文檔內容}For the full evaluation, training details and honesty statement, see the main repo model card.
The base model is licensed under the Gemma Terms of Use, and use of this quantized model is equally subject to those terms.
Gemma compliance statement
Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms
Note: the GGUF filenames keep the
xcloud-emb-*prefix unchanged, so that existing download scripts and ollama configurations do not break; the model itself is xVITA-Embed-300M-zhTW.
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