Small scale AI research: Expectations vs. Reality — A Stark Illustration of Limitations
The output from TinyLlama after extensive fine-tuning is, to put it bluntly, disheartening. While the model has grasped some basic structural elements of the desired output format, such as using <thinking>
and <output>
tags, the content itself is superficial, repetitive, and lacks coherence. The model struggles to maintain a logical flow of ideas and fails to close tags properly, resulting in an output that falls far short of the sophisticated reasoning I had hoped to achieve.
To fully illustrate the extent of this disappointment, let’s look at a sample from my training data and compare it to the model’s output:
Sample Training Data:
{
"Question": "What strategies can be employed to ensure successful integration of artificial intelligence systems into existing organizational infrastructure, while minimizing disruptions to existing workflows and employee routines?",
"Answer": "<thinking>\n\n**Sub-question 1: What are the key components of successful AI integration?**\n\n* Identify the organization's goals and objectives for AI integration\n* Determine the type of AI system being integrated (e.g., machine learning, natural language processing, robotic process automation)\n* Consider the existing infrastructure and systems that…