adpretko/armv8mac_to_riscv_qwen25coder_3p0b_full

Model Overview

This model, adpretko/armv8mac_to_riscv_qwen25coder_3p0b_full, is a specialized 3.1 billion parameter language model built upon the Qwen2.5-Coder-3B-Instruct architecture. It has been meticulously fine-tuned to address the specific challenge of translating code between different processor architectures.

Key Capabilities

• Cross-Architecture Code Translation: The model's core strength lies in its ability to convert code from ARMv8-A (armv8mac) to RISC-V. This is achieved through fine-tuning on a comprehensive set of armv8mac_to_riscv datasets.
• Instruction-Following: Inheriting capabilities from its base model, it is designed to follow instructions, which is crucial for targeted code generation and translation tasks.

Training Details

The model underwent a focused training regimen with the following hyperparameters:

• Learning Rate: 2e-05
• Batch Size: 1 (train), 8 (eval)
• Gradient Accumulation: 8 steps, resulting in an effective total train batch size of 8.
• Optimizer: AdamW with default betas and epsilon.
• Scheduler: Cosine learning rate scheduler with a 0.03 warmup ratio.
• Epochs: Trained for 0.5 epochs, indicating a highly targeted fine-tuning process on the specific datasets.

Intended Use Cases

This model is particularly well-suited for developers and researchers working on:

• Migrating existing ARMv8-A codebases to RISC-V platforms.
• Automating parts of the cross-compilation or code porting process.
• Research into architectural translation and code generation for embedded systems or custom hardware.