Sepolian/qwen2.5-0.5b-sft-openorca

Model Overview

Sepolian/qwen2.5-0.5b-sft-openorca is a 0.5 billion parameter causal language model, fine-tuned from the Qwen/Qwen2.5-0.5B base model. Its primary objective is to enhance instruction following and reasoning capabilities across diverse tasks. The model was trained on a carefully curated 300k sample subset of the OpenOrca dataset, emphasizing high-quality data for improved performance.

Key Capabilities

• Instruction Following: Designed to accurately interpret and execute user instructions.
• Reasoning: Exhibits improved reasoning abilities due to its training on a dataset focused on complex prompts and responses.
• English Language Processing: Optimized for tasks in the English language.
• Efficient Training: Achieved its fine-tuning in approximately 5 hours on NVIDIA L4 hardware, utilizing bf16 mixed precision and an AdamW optimizer.

Training and Data Quality

The model's training involved a rigorous data cleaning pipeline applied to the OpenOrca dataset. This pipeline included:

• Length & Refusal Filtering: Removing short or unhelpful responses.
• Language Filtering: Ensuring only English samples were used.
• Repetition Filtering: Eliminating excessive word or n-gram repetition.
• MinHash Deduplication: Removing near-duplicate samples to ensure data diversity.
• System Prompt Standardization: Unifying system prompts for consistent instruction interpretation.

Good For

• Instruction-tuned applications: Ideal for chatbots, virtual assistants, and other systems requiring precise instruction adherence.
• Reasoning tasks: Suitable for scenarios where logical deduction and coherent explanations are needed.
• Resource-constrained environments: Its 0.5B parameter size makes it a viable option for deployment where larger models might be impractical.