CodeFuse-CGM

codefuse-ai/CodeFuse-CGM is ranked 17,543 among these 8 peers; clawwork-ai/ClawWork is ranked 17,542.

CGM: Code Graph LLM Contents News Introduction Installation Examples Rewriter Retriever Reranker Reader Contributing Citation Join Us News 🔥🔥🔥 [2025/09/19] Our paper Code Graph Model (CGM): A Graph-Integrated Large Language Model for Repository-Level Software Engineering Tasks has been accepted to NeurIPS 2025! 🔥🔥🔥 [2025/01/15] We are pleased to announce the updated version of the CGM-72B-V1. The model further achieves a remarkable 44. 00% resolve rate on the SWE-Bench-Lite leaderboard. 🔥🔥🔥 [2024/12/28

] We are pleased to announce the updated version of the CGM-72B-V1. The model further achieves a remarkable 41. 67% resolve rate on the SWE-Bench-Lite leaderboard. 🔥🔥🔥 [2024/10/28] We are pleased to announce that CGM-72B achieves a remarkable 35. 67% resolve rate on the SWE-Bench-Lite leaderboard. 🔥🔥🔥 [2024/10/28] We released **CGM**, mainly for repository-level coding tasks. 📜 **Paper**: Code Graph Model (CGM): A Graph-Integrated Large Language Model for Repository-Level Software

Engineering Tasks 🤖 **Model**: codefuse-ai/CodeFuse-CGM-72B 📊 **Data**: codefuse-ai/CodeGraph Introduction We propose a graph-based framework CGM for real-world SE tasks. Before CGM starts its work, we construct a repository-level code graph to better represent the repository context and its structure by Code Graph Generator. Inspired by the Retrieval-Augmented Generation (RAG) approach, CGM framework is designed as a chain structure consisting of four atomic nodes, termed as R4 (Rewriter, Retriever, Reranker, and Reader) chain

for this scenario. Given an issue, the initial input to the CGM framework includes the issue description and the corresponding code graph. Rewriter will first rewrite the original issue by extracting keywords and generating relevant queries for code graph. Then a heuristic code subgraph is retrieved through Retriever based on the matching anchor nodes from rewriter output. Given that the resulting subgraph provides a relatively broad context necessary for reference, we need a Reranker to identify the files most likely

to be modified as a further hint. Subsequently, both the retrieved subgraph and the identified files are input into a trainable, graph-based Reader to generate the corresponding code patch. Framework Highlights :white_check_mark: **Code Graph**: Train models on multiple tasks while maintaining a balance between them. The models can even generalize to new, previously unseen tasks. :white_check_mark: **Multi-framework**: It provides support for both Accelerate (with Deepspeed and FSDP) :white_check_mark: **Efficient fine-tuning**: It

supports LoRA, QLoRA as well as Full-parameters training, enabling fine-tuning of large models with minimal resources. The training speed meets the demands of almost all fine-tuning scenarios. 8+ pip Required Packages Examples The following chart illustrates the whole processing pipeline of R3. Pre-process for Retriever: Generate Node Embedding Before Retriever, we need to embed all the nodes in Code Graph into embeddings Queries generated by Rewriter, into embeddings by CGE-large Requirements for CGE-large Rewriter

Given issues meta data, execute the following scripts to generate Rewriter results (both Inferer and Extractor). Use function and in