CodeModernBERT-Crow-v1-Pre

Model Description

CodeModernBERT-Crow-v1-Pre is a pretrained language model based on the ModernBERT architecture, specifically adapted for source code and docstring style natural language. It supports multiple programming languages and is trained using large-scale code datasets curated from open-source repositories.

• License: Apache-2.0

• Supported Languages: Python, JavaScript, TypeScript, Java, Go, Rust, PHP, Ruby, C++, C, SQL

• Datasets:

  • bigcode/starcoderdata
  • bigcode/starcoder2data-extras

• Pipeline tag: fill-mask

This model is a pretraining checkpoint, designed for further fine-tuning on downstream tasks such as semantic code search, bug detection, or code summarization.

---

Training Objective

The model was pretrained on large-scale multilingual code corpora with the following goals:

• Learn robust code representations across multiple programming languages.
• Capture semantic relations between code tokens and natural language descriptions.
• Provide a strong initialization point for fine-tuning on code-related downstream tasks.

---

Tokenizer

A custom BPE tokenizer was trained for code and docstrings.

• Vocabulary size: 50,368

• Special tokens: Standard Hugging Face special tokens + custom tokens for code/document structure.

• Training process:

  • Up to 1M examples per dataset.
  • Each example truncated to 10,000 characters.
  • Trained with files from multiple datasets (see above).

---

Architecture

• Base: ModernBERT
• Hidden size: 768
• Number of layers: 12
• Attention heads: 12
• Intermediate size: 3072
• Max sequence length: 8192 (during training, inputs were limited to 1024 tokens)
• RoPE positional encoding: supported

---

Usage

from transformers import AutoTokenizer, AutoModelForMaskedLM

tokenizer = AutoTokenizer.from_pretrained("Shuu12121/CodeModernBERT-Crow-v1-Pre")
model = AutoModelForMaskedLM.from_pretrained("Shuu12121/CodeModernBERT-Crow-v1-Pre")

inputs = tokenizer("def add(a, b): return a + b", return_tensors="pt")
outputs = model(**inputs)

The model can be fine-tuned for:

• Code search (query ↔ code retrieval)
• Code clone detection
• Code summarization (docstring prediction)
• Bug detection and repair (masked language modeling or cloze-style)

---

Limitations

• The model is not optimized for direct code generation.
• Pretraining does not guarantee correctness of code execution.
• Fine-tuning is recommended for specific downstream applications.

---

Intended Use

• Research in software engineering and natural language processing for code.
• Educational exploration of pretrained models for code tasks.
• Baseline for continued pretraining or fine-tuning.

---