Shuu12121 / NightOwl-CodeEmbedding

NightOwl Code Embedding

NightOwl-CodeEmbedding is a 768-dimensional dense embedding model for code search, code-to-code retrieval, code editing retrieval, and technical QA.

It is based on Shuu12121/NightOwl and fine-tuned with hard negatives from code search, code editing, and CoIR retrieval datasets. The model uses CLS pooling and cosine similarity.

No query or document prefix is required.

Usage

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("Shuu12121/NightOwl-CodeEmbedding")

queries = ["Python function that sorts a list in descending order"]
documents = [
    "def sort_desc(values): return sorted(values, reverse=True)",
    "def average(values): return sum(values) / len(values)",
]

query_embeddings = model.encode(queries, normalize_embeddings=True)
document_embeddings = model.encode(documents, normalize_embeddings=True)
scores = query_embeddings @ document_embeddings.T
print(scores)

Model Details

MTEB Results

Evaluated locally with MTEB 2.5.1 on MTEB(Code, v1). Scores are NDCG@10. Multi-subset task scores are macro averages.

Several benchmark families were also represented in the hard-negative fine-tuning data. Treat these scores as an in-domain evaluation rather than a strict zero-shot result.

Training

The model was fine-tuned without task prefixes using bidirectional contrastive learning and hard negatives from:

• Shuu12121/codeedit_hard_negative_datasets_kd
• Shuu12121/owl_code_search_hard_negative_datasets_V2_kd
• Shuu12121/coir_hard_negative_datasets_v2_kd

The training loss was CachedMultipleNegativesRankingLoss .