aiXcoder/aiXapply-4B-SFT
aiXcoder/aiXapply-4B-SFT is a 4 billion parameter supervised fine-tuned model developed by aiXcoder, specifically designed for the 'Full-File Apply' code integration task. It takes an original file and an update snippet to generate a complete, updated file, preserving unchanged code. This model achieves 94.4% average equivalence accuracy on its main benchmark and demonstrates strong long-context structural preservation, making it ideal for reliable code integration in IDE workflows.
Loading preview...
aiXapply-4B-SFT: Specialized Full-File Code Integration
aiXapply-4B-SFT is a 4 billion parameter model from aiXcoder, specifically supervised fine-tuned for the Full-File Apply task. This involves taking an original code file and a localized update snippet, then generating the complete, updated file while ensuring all unrelated code remains unchanged. It is designed to be a robust backend for IDE workflows, addressing the challenge of reliably applying code snippets.
Key Capabilities and Features
- High Accuracy: Achieves 94.4% average equivalence accuracy on a 1,637-sample benchmark, outperforming models like DeepSeek-V3.2 (91.6%) and close to Qwen3.5-397B-A17B (94.8%).
- Long-Context Fidelity: Demonstrates superior structural preservation in long-context scenarios compared to its RL-aligned counterpart.
- Fast Generation: When combined with n-gram speculative decoding, it can achieve 1.06s average latency and 2692 tokens/s on an A100 GPU.
- Deployment-Ready: Can be served via an OpenAI-compatible endpoint and integrated as a dedicated
applymodel in tools like Continue. - Comprehensive Pipeline: The release includes data generation, training scripts (SFT and RL), inference, evaluation, and error classification tools.
- Task Definition: Strictly adheres to requirements for complete output, no side effects outside the edit, and proper placeholder expansion.
Use Cases and Differentiators
This model is specifically optimized for the precise and reliable integration of code snippets into full files, a critical step in modern coding assistant workflows. Unlike general-purpose LLMs that might struggle with exact file reconstruction or introduce side effects, aiXapply-4B-SFT is purpose-built for this deterministic task. It offers a significant improvement over traditional methods like unified diffs or search-and-replace for code integration accuracy and efficiency, making it a strong choice for developers needing a dedicated, high-performance code application engine.