MatthewsO3/GraphCode-CErl-codesearch

GraphCode-CErl — Semantic Code Search for Erlang & C++

Fine-tuned GraphCodeBERT for semantic code search over Erlang and C++ codebases. Given a natural language query, the model retrieves the most semantically relevant functions from an indexed repository.

Model Description

This is a bi-encoder trained with contrastive learning. It encodes both natural language queries and code snippets into a shared embedding space, enabling efficient cosine-similarity-based retrieval at search time.

• —Base model: microsoft/graphcodebert-base
• —Architecture: GraphCodeBERT encoder with mean pooling + L2 normalization (no LM head)
• —Languages trained on: Erlang, C++
• —Task: Semantic code search / function retrieval

Architecture detail

The model wraps the GraphCodeBERT encoder in a lightweight CodeSearchModel:

# Mean pooling over all token positions (not CLS)
def mean_pooling(last_hidden_state, attention_mask):
    mask = attention_mask.unsqueeze(-1).expand(last_hidden_state.size()).float()
    return torch.sum(last_hidden_state * mask, 1) / torch.clamp(mask.sum(1), min=1e-9)

Embeddings are L2-normalized, so retrieval is a plain dot product (equivalent to cosine similarity).

Training

Data

Training triplets were constructed from two sources:

Each record is a (code, good_docstring, bad1_docstring, bad2_docstring) tuple. Negatives were mined as follows:

• —60% hard negatives — BM25-retrieved docstrings that are lexically similar to the positive but semantically wrong (top-20 BM25 candidates, sampled randomly)
• —30% cross-language negatives — docstrings sampled from the opposite language to discourage language-specific shortcuts
• —10% random negatives — uniform random docstrings as easy negatives

Loss

Temperature-scaled cross-entropy over augmented scores. For each batch the score matrix is extended with both negatives:

augmented_scores = [good_scores | bad1_scores | bad2_scores]
loss = CrossEntropyLoss(augmented_scores / τ, diagonal_labels)

where τ = 0.05.

Hyperparameters

Training curve

Training ran for all 10 epochs without triggering early stopping (patience = 3). Best model saved at epoch 10.

Usage

This model is intended to be used with `code_search.py`, a unified indexing and search tool included in the repository.

Quick start

git clone https://github.com/MatthewsO3/GraphCode-CErl-base
cd "GraphCode-CErl-base/Code Search/Evaluation"
python setup.py          # creates .venv, installs deps, builds erlang.so
source .venv/bin/activate

# Index a repository (auto-discovers Erlang + C++ + Python)
python code_search.py index \
    --repo /path/to/your/repo \
    --model MatthewsO3/GraphCode-CErl-codesearch \
    --output corpus.jsonl \
    --index corpus_index.pt

# Search interactively
python code_search.py search \
    --model MatthewsO3/GraphCode-CErl-codesearch \
    --jsonl corpus.jsonl \
    --index corpus_index.pt \
    --top 5

Language-specific flags are also available and can be combined freely:

# Erlang only
python code_search.py index --erlang /path/to/erl_repo ...

# C++ only
python code_search.py index --cpp /path/to/cpp_repo ...

# Explicit mix
python code_search.py index --erlang /path/erl --cpp /path/cpp --python /path/py ...

Using the model directly

from transformers import AutoTokenizer, AutoModel
import torch

tokenizer = AutoTokenizer.from_pretrained("microsoft/graphcodebert-base")
model = AutoModel.from_pretrained("MatthewsO3/GraphCode-CErl-codesearch")
model.eval()

def encode(texts):
    enc = tokenizer(texts, return_tensors="pt", truncation=True,
                    padding=True, max_length=256)
    with torch.no_grad():
        out = model(**enc)
    # Mean pooling
    mask = enc["attention_mask"].unsqueeze(-1).float()
    emb = (out.last_hidden_state * mask).sum(1) / mask.sum(1).clamp(min=1e-9)
    return emb / emb.norm(dim=1, keepdim=True)

query = encode(["handle TCP connection timeout"])
code  = encode(["handle_timeout(Socket, State) -> gen_tcp:close(Socket), {stop, timeout, State}."])

score = (query @ code.T).item()
print(f"Similarity: {score:.4f}")

Note: The tokenizer is loaded from microsoft/graphcodebert-base since it is identical to the fine-tuned model's tokenizer and avoids a redundant download.

Supported Languages

Note: Python indexing is supported by code_search.py but the model was not trained on Python data. Results for Python queries may be less accurate.

Limitations

• —Not trained on Python — cross-language transfer to Python is best-effort
• —The Erlang training set is private and not released
• —Functions without docstrings or comments are embedded on code tokens alone, which may reduce retrieval accuracy for ambiguous natural language queries
• —Running on CPU is fully supported but slow for large corpora at index-build time; a GPU is recommended

Repository

Training code, indexing tool, and setup scripts are available at: github.com/MatthewsO3/GraphCode-CErl-base

Citation

If you use this model, please cite the original GraphCodeBERT paper:

@inproceedings{guo2021graphcodebert,
  title     = {GraphCodeBERT: Pre-training Code Representations with Data Flow},
  author    = {Guo, Daya and Ren, Shuo and Lu, Shuai and Feng, Zhangyin and Tang, Duyu
               and Liu, Shujie and Zhou, Long and Duan, Nan and Svyatkovskiy, Alexey
               and Fu, Shengyu and others},
  booktitle = {International Conference on Learning Representations},
  year      = {2021}
}