Bounded Brainstorming Finetune

Bounded-Brainstorming-Finetune addresses the limitations of general-purpose LLMs in generating novel content that adheres to strict formatting and exclusion criteria. By providing a task description, a list of existing items, and a target format, the model is trained to produce a specific number of new entries that remain in-domain while avoiding duplication of provided priors. This approach aims to solve common issues such as model collapse, format degradation, and repetitive output.

The project utilizes a systematic approach involving synthetic data generation, embedding-based filtering to ensure novelty, and supervised fine-tuning of small, open-weight instruction models. It includes a comprehensive evaluation framework using metrics like format adherence, intra-set diversity, and cosine similarity against prior entries. This tooling is particularly useful for workflows requiring structured creative generation, such as ideating podcast prompts, defining ASR training sets, or drafting professional career trajectories.