README.md

metadata

language:
  - en
  - code
tags:
  - python
  - code-search
  - code-embeddings
  - sentence-transformers
  - unixcoder
  - matryoshka
  - onnx
  - mteb
license: cc-by-nc-4.0
pipeline_tag: feature-extraction
library_name: sentence-transformers
metrics:
  - mrr
  - ndcg
  - recall
model-index:
  - name: Code-Embed
    results:
      - task:
          type: feature-extraction
          name: Code Search
        dataset:
          name: CodeSearchNet (Python)
          type: code_search_net
          config: python
          split: test
        metrics:
          - type: mrr@10
            value: 0.7514
            name: MRR@10 (100-pool)
          - type: mrr@10
            value: 0.8241
            name: MRR@10 (15k-corpus)
          - type: recall@10
            value: 0.923
            name: Recall@10 (15k-corpus)
      - task:
          type: feature-extraction
          name: Code Question Answering
        dataset:
          name: CoSQA
          type: cosqa
          split: validation
        metrics:
          - type: mrr@10
            value: 0.7509
            name: MRR@10
          - type: recall@10
            value: 0.86
            name: Recall@10

💻 code-embed

CodeEmbed is a lightweight, high-performance code embedding model (≤130M parameters) developed by Divakant Pandey. Specialized for semantic code search, code retrieval, and repository navigation with 8,192 token context and Matryoshka representation learning.

🚀 Key Features

• Author / Maintainer: Divakant Pandey
• Backbone: UniXcoder bi-encoder architecture with a 2-layer MLP projection head and unit-sphere L2 normalization.
• 3-Stage Curriculum:
  1. SimCSE Unsupervised Contrastive Learning (uniformity on hypersphere).
  2. Supervised Fine-Tuning (MNRL with BM25 lexical hard negatives).
  3. Matryoshka Representation Learning (MRL) (nested dimensions: 64, 128, 256, 384, 768, 1536).
• Fast CPU Inference: ~20.8 ms per code snippet on standard laptop CPU.
• Production-Ready: Exportable to ONNX with dynamic axes and dynamic INT8 quantization.

📊 Benchmark Results

Benchmark	Metric	Value	Gate Status
CodeSearchNet (Python)	MRR@10	0.7514	PASS (≥ 0.65)
CoSQA (Code Search QA)	MRR@10	0.7509	PASS (≥ 0.50)
Hypersphere Uniformity	Uniformity	-1.9758	PASS (≤ -1.0)
Single-Snippet Latency	CPU Latency	20.8 ms	PASS (< 100 ms)

💻 Quickstart Usage

With Sentence-Transformers:

from sentence_transformers import SentenceTransformer

# Automatically downloads from Hugging Face Hub
model = SentenceTransformer("divakantpandey/code-embed")

queries = ["binary search in sorted list", "read json file"]
snippets = [
    "def binary_search(arr, target): ...",
    "import json\ndef load_json(path): return json.load(open(path))"
]

query_embeddings = model.encode(queries, normalize_embeddings=True)
code_embeddings = model.encode(snippets, normalize_embeddings=True)

# Compute cosine similarity
similarities = query_embeddings @ code_embeddings.T
print(similarities)

📜 License & Commercial Inquiries

This model and its weights are released under the Creative Commons Attribution-NonCommercial 4.0 International (CC-BY-NC 4.0) license.

• Non-Commercial Use: Free for academic research, education, benchmarking, and non-commercial projects with attribution.
• Attribution: Please cite Divakant Pandey and link to https://github.com/divakantpandey/code-embed.
• Commercial Licensing: To use this model in commercial applications, proprietary software, enterprise APIs, or SaaS products, please contact:
  • Maintainer: Divakant Pandey
  • Email: divakant01@gmail.com
  • GitHub: @divakantpandey