Technical Overview
Conan‑embedding‑v1 is a sentence‑level embedding model released by TencentBAC. Built on the sentence‑transformers framework and powered by PyTorch, it converts Chinese sentences (and multilingual text when needed) into dense vector representations that capture semantic similarity. The model is distributed as safafetensors weights, making it lightweight and fast to load.
Key capabilities include:
- High‑quality semantic similarity scoring for Chinese (zh) corpora.
- Support for downstream tasks such as clustering, retrieval, reranking, and pairwise classification.
- Compatibility with the sentence‑transformers pipeline, allowing zero‑shot use in many NLP applications.
Architecture highlights – The model adopts a BERT‑style encoder (the exact backbone is not disclosed, but the tag list includes bert and sentence‑transformers) and adds a pooling layer that generates a fixed‑size embedding (typically 768 dimensions). The use of torch‑safetensors ensures memory‑efficient loading, while the training regime follows the standard Masked Language Modeling + Sentence‑level Contrastive Learning paradigm common in modern embedding models.
Intended use cases revolve around any scenario where you need a numeric representation of a sentence for similarity comparison: semantic search over Chinese documents, duplicate detection, clustering of user reviews, or as a feature extractor for downstream classifiers. Because the model is optimized for the Chinese language, it excels on tasks where high‑quality Chinese textual understanding is required.
Benchmark Performance
The model’s performance is reported on the MTEB (Massive Text Embedding Benchmark) suite, which aggregates a wide range of evaluation tasks such as semantic textual similarity (STS), classification, clustering, reranking, and retrieval. These benchmarks are crucial for embedding models because they directly measure how well the vectors preserve semantic relationships across diverse domains.
Key results (average scores) include:
- STS (AFQMC & ATEC) – Cosine‑based Pearson scores around 56.6 % and 56.6 % respectively, with Spearman scores in the low‑60 % range.
- STS (BQ) – Cosine Pearson 72.6 % and Spearman 74.5 %, indicating strong performance on Chinese news‑style sentences.
- Classification (Amazon Reviews – zh) – Accuracy 50.3 % and F1 46.9 % on a multi‑class sentiment task.
- Clustering – V‑measure of 60.6 % for CLSClusteringP2P and 52.6 % for CLSClusteringS2S.
- Reranking (CMedQAv1 & CMedQAv2) – MAP scores above 89 % and MRR above 91 % on medical QA reranking.
- Retrieval (CmedqaRetrieval) – MAP@10 40.7 %, MRR@10 49.8 %, and NDCG@10 47.6 % on a large‑scale medical QA corpus.
- Pairwise Classification (CMNLI) – Cosine‑accuracy 85.9 % and AP 92.5 %.
Compared with other Chinese sentence‑embedding models (e.g., nghuyong/ernie‑3.0‑base‑zh or shibing624/text2vec‑baidu‑qa‑2020), Conan‑embedding‑v1 shows competitive STS scores while excelling in medical‑domain reranking and retrieval, making it a strong candidate for domain‑specific applications.
Hardware Requirements
The model’s checkpoint is stored in safafetensors format and occupies roughly 1.2 GB on disk. For inference, the typical VRAM footprint is about 2–3 GB for a batch size of 1 (single‑sentence inference) on a modern GPU. Larger batch sizes or mixed‑precision (FP16) can reduce memory usage further.
- GPU – Any NVIDIA GPU with at least 6 GB VRAM (e.g., RTX 3060, GTX 1660 Ti) will run the model comfortably. For high‑throughput batch processing, a 12 GB or higher GPU (RTX 3080, A100) is recommended.
- CPU – A recent multi‑core CPU (Intel i5‑10600K or AMD Ryzen 5 5600X) is sufficient for tokenization and data loading. The model does not require GPU‑only execution; CPU inference is possible but will be slower (≈3–5× latency).
- Storage – The model files (weights, config, tokenizer) total roughly 1.5 GB. SSD storage is advised for fast loading.
- Performance – On an RTX 3060 (FP16), single‑sentence inference averages ≈15 ms. Batch inference of 32 sentences drops to ≈5 ms per sentence due to parallelism.
Use Cases
Conan‑embedding‑v1 shines in any workflow that requires dense semantic representations of Chinese text. Typical applications include:
- Semantic Search – Indexing Chinese documents (news, legal texts, FAQs) and retrieving the most relevant passages via cosine similarity.
- Duplicate Detection – Identifying near‑duplicate user reviews, forum posts, or product descriptions.
- Clustering & Topic Modeling – Grouping large corpora of Chinese sentences into coherent clusters for analytics.
- Reranking in QA Systems – Re‑ordering candidate answers in medical or technical Q&A platforms, as demonstrated by the CMedQAv1/2 benchmarks.
- Feature Extraction for Classification – Feeding sentence embeddings into downstream classifiers (sentiment analysis, intent detection).
Industries that benefit most are e‑commerce (review analysis), healthcare (medical QA), finance (document retrieval), and government (policy document clustering). The model can be integrated via the sentence‑transformers Python library, Hugging Face transformers pipelines, or exported to ONNX for deployment in production services.
Training Details
While the README does not disclose the full training pipeline, the model follows the standard sentence‑transformers training paradigm. It was likely pre‑trained on a large Chinese corpus using masked language modeling, then fine‑tuned with a contrastive loss on the MTEB datasets (e.g., AFQMC, ATEC, BQ) to align sentence vectors.
Datasets – The model was evaluated on several Chinese benchmark datasets, suggesting that training data included:
- AFQMC (Chinese sentence pair similarity)
- ATEC (Chinese semantic textual similarity)
- BQ (Chinese news sentences)
- Medical QA corpora (CMedQAv1/2)
- Amazon multi‑language reviews (zh)
Compute – Training a BERT‑style encoder with contrastive objectives typically requires 8–16 A100 GPUs for 2–3 days, depending on batch size and dataset size. The final checkpoint size (~1.2 GB) indicates a base‑size model (≈110 M parameters).
Fine‑tuning – Users can further adapt the model to domain‑specific data via the SentenceTransformer API, employing a small learning rate (1e‑5 – 5e‑5) and a contrastive or triplet loss. The model’s safetensors format makes it straightforward to load and continue training in PyTorch.
Licensing Information
The repository lists the license as license:cc‑by‑nc‑4.0 (Creative Commons Attribution‑NonCommercial 4.0). However, the overall License field is marked “unknown”. In practice, the CC‑BY‑NC‑4.0 tag means you may use, share, and adapt the model for non‑commercial purposes provided you give appropriate credit to TencentBAC.
Because the license is non‑commercial, commercial deployment (e.g., SaaS products, paid APIs) is not permitted without obtaining a separate commercial agreement from the author. You must also retain the attribution notice in any derivative work and include a link to the original model card.
If you plan to integrate the model into a proprietary system, you should contact TencentBAC to negotiate a commercial license or seek an alternative model with a permissive license (e.g., MIT, Apache 2.0). Failure to comply with the non‑commercial clause could result in copyright infringement.