Kiel-2.5-Poly

Kiel-2.5-Poly is a state-of-the-art, high-performance multilingual text and code embedding model developed by KielTech. Built upon the robust BAAI/bge-m3 architecture and fine-tuned using Unsloth-accelerated QLoRA, this model maps sentences, code blocks, and security vulnerability briefs into a dense 768-dimensional vector space.

It is specifically optimized for cross-lingual semantic similarity, retrieval-augmented generation (RAG), automated code auditing, and security-based reasoning.

---

Model Details

• Creator: KielTech
• Base Model: BAAI/bge-m3
• Model Type: Sentence Transformer / Multilingual Dense Embedding Backbone
• Output Dimensionality: 768 dimensions
• Maximum Context Length: Up to 8,192 tokens
• Similarity Function: Cosine Similarity
• Supported Modalities: Multilingual Text, Python Code Snippets, and Security CoT Corpora

---

Training & Fine-Tuning Overview

Kiel-2.5-Poly was fine-tuned on a balanced 40,000-sample multi-domain dataset consisting of:

1. QQP Triplets: Paraphrase identification and semantic similarity mapping.
2. All-NLI Corpus: Semantic entailment and logical reasoning patterns.
3. CodeSearchNet: Python code-documentation and syntax alignment.
4. Curated Security CoT: Multi-step vulnerability analysis and mitigation reasoning traces (JWT validation, SQLi, XSS, and path traversal).

Optimization was performed using Multiple Negatives Ranking Loss combined with memory-efficient 4-bit QLoRA via Unsloth.

---

Direct Usage

First, ensure you have the sentence-transformers library installed:

pip install -U sentence-transformers

Then load your model and run inference:

from sentence_transformers import SentenceTransformer

# Load Kiel-2.5-Poly from the Hugging Face Hub
model = SentenceTransformer("kiel2/Kiel-2.5-Poly")

# Run inference on multilingual or technical text
sentences = [
    "How to securely validate and decode JWT tokens in Python backend services",
    "Step 1: Inspect the token header. Step 2: Enforce cryptographic signature verification against a secure key.",
    "print('Hello from KielTech code pipeline')"
]

embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 768]

# Compute similarity scores
similarities = model.similarity(embeddings, embeddings)
print(similarities)
Citation
If you use Kiel-2.5-Poly in your research, applications, or commercial pipelines, please cite it as follows:

Code snippet
@misc{kiel2026kiel25poly,
  title={Kiel-2.5-Poly: Advanced Multilingual, Code, and Security Embedding Model},
  author={KielTech},
  year={2026},
  publisher={Hugging Face Hub},
  howpublished={\url{[https://huggingface.co/kiel2/Kiel-2.5-Poly](https://huggingface.co/kiel2/Kiel-2.5-Poly)}}
}