Xieral Code Gen 3B

Xieral Code Gen 3B is a decoder-only language model with 2.7 billion parameters. Developed from the Xieral-Code-Gen-3b, this model is designed specifically for code generation and software engineering tasks.

Model Overview

Architecture: Decoder-only language model
Parameters: 2.7 billion
Training Data: Combination of publicly available and synthetic datasets
Optimization: Direct Preference Optimization (DPO)
Fine-tuning: General code/software engineering conversations, SQL query generation, and discussion

Performance

Xieral Code Gen 3B has demonstrated competitive performance compared to other models of similar size:

MultiPL-E Metrics: Evaluated across various programming languages using the BigCode Evaluation Harness.
MT Bench: Shows strong results on code-related tasks.

Usage

This model is well-suited for:

General code/software engineering conversations
SQL query generation and discussion

Requirements

To run Xieral Code Gen 3B locally, you will need:

VRAM: 8GB+ (Graphics card with sufficient VRAM)
Dependencies: Ensure you have the necessary libraries and environment set up to run the model.

Installation

To install the required dependencies, use:

pip install -r requirements.txt