Exploring Open-Source Large Language Models for InkMind: Choosing a Better Poetry Continuation Model

When selecting an open-source large language model for InkMind's poetry continuation feature, several models were evaluated: GPT-Neo-125M, Pythia-160M, and Qwen3-1.7B-Base. GPT-Neo-125M is a smaller model developed by EleutherAI, with 125 million parameters and a context length of 2,048 tokens. It is licensed under MIT and can be used with Hugging Face's Transformers library without a paid API.

While GPT-Neo-125M is lightweight and easy to work with, its capacity may be limiting for complex poetry generation. Pythia-160M, also from EleutherAI, contains 160 million parameters and follows a decoder-only Transformer architecture. Trained on The Pile dataset, Pythia is licensed under Apache 2.0 and offers more research-oriented options due to multiple model checkpoints.

However, it may not be specifically optimized for poetry generation, so experimental evaluation would be necessary. Lastly, Qwen3-1.7B-Base was developed by Alibaba's Qwen team. This larger model consists of 1.7 billion parameters with a 32,768 token context window, making it suitable for capturing complex language patterns across longer passages.

The model's Apache 2.0 license also permits its use with Hugging Face's Transformers library. Despite its impressive capacity, Qwen3-1.7B-Base requires more computational resources, which may not be ideal for Google Colab users with memory constraints. Considering the project's goals, GPT-Neo-125M's lightweight nature, Pythia-160M's flexibility for fine-tuning, and Qwen3-1.7B-Base's larger context window each have their merits.

Further experimentation with tokenization, embeddings, self-attention, temperature, top-k, and top-p settings on Google Colab will be necessary to determine the most effective model for InkMind's poetry generation.