all-MiniLM-L6-v2-code-search-512

Version: v1.0
Release Date: 2026-01-04
Base Model: https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2

---

Overview

• all-MiniLM-L6-v2-code-search-512 is a lightweight, high-accuracy sentence-transformers fine-tuned specifically for semantic code search and code embeddings.
• It maps natural language queries and source code into a shared embedding space, enabling fast and accurate retrieval of relevant code snippets across multiple programming languages.
• On CodeSearchNet, this model achieves 90.3% Accuracy@1, significantly outperforming the base all-MiniLM-L6-v2 model while keeping the same 33M parameter footprint.

Key Features

• Fine-tuned on 1.29M+ code–documentation pairs from CodeSearchNet
• Supports 512-token context length (2× longer than the base model)
• Multi-language code support (Python, Java, JavaScript, PHP, Ruby, Go)
• 33M parameters for fast inference
• Pre-configured with max_seq_length=512 by default

Intended Use

Recommended use cases:

• Semantic code search
• Natural language to code retrieval
• Code–documentation matching
• Code similarity and clustering
• Developer tools and IDE integrations

Supported languages:

• Python
• Java
• JavaScript
• PHP
• Ruby
• Go
• Other languages included in CodeSearchNet

Quick Start

from sentence_transformers import SentenceTransformer

model = SentenceTransformer('isuruwijesiri/all-MiniLM-L6-v2-code-search-512')

# Encode code
code_embedding = model.encode("def hello_world():\n    print('Hello, World!')")

# Encode natural language description
query_embedding = model.encode("function that prints hello world")

# Calculate similarity
from sklearn.metrics.pairwise import cosine_similarity
similarity = cosine_similarity([query_embedding], [code_embedding])[0][0]
print(f"Similarity: {similarity:.4f}")

Quick Start - JS

Use this model in JavaScript/TypeScript with Transformers.js - works in Node.js and browsers!

Installation:

npm install @xenova/transformers

Usage:

import {{ pipeline }} from '@xenova/transformers';

// Load the model (auto-downloads and caches)
const extractor = await pipeline('feature-extraction', 'isuruwijesiri/all-MiniLM-L6-v2-code-search-512', {
  quantized: false  // false = model.onnx (86MB), true = model_quantized.onnx (22MB)
});

// Generate embeddings
const output = await extractor('def add(a, b): return a + b', {{
  pooling: 'mean',
  normalize: true
}});

// Use the embedding
const embedding = Array.from(output.data);
console.log(embedding); // [384] dimensional vector

Code Search Example

from sentence_transformers import SentenceTransformer, util

# Load model
model = SentenceTransformer('isuruwijesiri/all-MiniLM-L6-v2-code-search-512')

# Your code snippets
code_snippets = [
    "def calculate_sum(a, b):\n    return a + b",
    "def find_max(numbers):\n    return max(numbers)",
    "class User:\n    def __init__(self, name):\n        self.name = name"
]

# Natural language query
query = "function to add two numbers"

# Encode
query_emb = model.encode(query, convert_to_tensor=True)
code_embs = model.encode(code_snippets, convert_to_tensor=True)

# Find most similar
similarities = util.cos_sim(query_emb, code_embs)[0]
best_match_idx = similarities.argmax().item()

print(f"Best match: {code_snippets[best_match_idx]}")
print(f"Similarity: {similarities[best_match_idx]:.4f}")

Code Search Example - JS

import {{ pipeline }} from '@xenova/transformers';

const extractor = await pipeline('feature-extraction', 'isuruwijesiri/all-MiniLM-L6-v2-code-search-512', {
  quantized: false  // false = model.onnx (86MB), true = model_quantized.onnx (22MB)
});

// Your code snippets
const codeSnippets = [
  "def add(a, b): return a + b",
  "def multiply(x, y): return x * y",
  "class User: pass"
];

// Search query
const query = "function to add two numbers";

// Get embeddings
const queryEmb = await extractor(query, {{ pooling: 'mean', normalize: true }});
const codeEmbs = await Promise.all(
  codeSnippets.map(code => extractor(code, {{ pooling: 'mean', normalize: true }}))
);

// Calculate similarities (dot product of normalized vectors = cosine similarity)
const similarities = codeEmbs.map(codeEmb => 
  Array.from(queryEmb.data).reduce((sum, val, i) => 
    sum + val * codeEmb.data[i], 0
  )
);

console.log(similarities); // [0.87, 0.23, 0.15] - first one is most similar!

Performance

Evaluated on the CodeSearchNet validation set (2,000 samples).

Metric	Score
MRR@10	0.9259
Accuracy@1	0.9030
Accuracy@3	0.9440
Accuracy@5	0.9540
Accuracy@10	0.9705
Recall@1	0.9030
Recall@5	0.9540
Recall@10	0.9705
NDCG@10	0.9367
MAP@100	0.9269

This model significantly outperforms the base all-MiniLM-L6-v2 on code search tasks:

Metric	Base Model	Fine-tuned	Improvement
MRR@10	0.7759	0.9259	+0.1500 (+19.3%)
Accuracy@1	0.7175	0.9030	+0.1855 (+25.9%)
Accuracy@5	0.8555	0.9540	+0.0985 (+11.5%)
Recall@10	0.8840	0.9705	+0.0865 (+9.8%)
NDCG@10	0.8022	0.9367	+0.1344 (+16.8%)

Context Length Configuration

This model was trained with a maximum sequence length of 512 tokens. By default, max_seq_length is already set to 512. No manual configuration is required.

model.max_seq_length = 256  # Faster inference for short code
model.max_seq_length = 512  # Default, best accuracy (recommended)

Training Details

Dataset: CodeSearchNet
Training samples: 1,294,017
Evaluation samples: 2,000
Epochs: 3
Batch size: 192
Loss function: MultipleNegativesRankingLoss

Limitations

• Inputs longer than 512 tokens are truncated
• Not suitable for code generation or completion
• Optimized for semantic similarity, not syntactic correctness
• Performance may vary across programming languages

---

Citation

If you use this model, please cite:

@misc{all_MiniLM_L6_v2_code_search_512,
  author = {isuruwijesiri},
  title = {all-MiniLM-L6-v2-code-search-512},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/isuruwijesiri/all-MiniLM-L6-v2-code-search-512}
}

---

License: Apache 2.0