Nondzu/Mistral-7B-Instruct-v0.2-code-ft Overview
This model is an instruction-tuned variant of the Mistral-7B-Instruct-v0.2 base model, developed by Nondzu, with a strong focus on enhancing coding assistance and co-pilot capabilities. It incorporates a 7 billion parameter architecture and supports an extended sequence length of 16384 tokens through sample packing.
Key Capabilities & Features
- Code-centric Fine-tuning: Specifically trained on the
ajibawa-2023/Code-74k-ShareGPT dataset to improve performance in code generation, completion, and understanding tasks. - Base Model Enhancements: Builds upon the robust Mistral-7B-Instruct-v0.2, leveraging its strong foundational language understanding.
- Optimized Training: Utilizes QLoRA adapter,
adamw_bnb_8bit optimizer, bf16 training, gradient checkpointing, and flash attention for efficient and effective fine-tuning. - ChatML Prompt Template: Designed to work with the ChatML prompt format for structured conversational interactions.
- Quantized Versions Available: Links to various quantized versions (EXL2, GGUF, AWQ, GPTQ) are provided for broader deployment flexibility.
Performance & Use Cases
- Eval Plus Score: Achieves a score of 0.421 on Eval Plus, indicating its proficiency in coding benchmarks.
- Ideal for: Developers seeking an open-source model for code generation, debugging assistance, and integrating co-pilot features into applications. Its specialized training makes it a strong candidate for programming-related tasks.