Overview

This model implements a diverse beam search strategy, a variant of the standard beam search algorithm, to generate more varied and less repetitive text outputs. It is built upon the Qwen/Qwen3-0.6B base model, a 0.8 billion parameter decoder-only transformer. The core idea is to divide the total number of beams into groups and apply a diversity penalty to discourage similar sequences within these groups, thereby increasing the overall diversity of the generated candidates.

Key Capabilities

• Enhanced Output Diversity: Produces a broader range of distinct text sequences compared to traditional beam search.
• Configurable Parameters: Allows fine-tuning of generation behavior through num_beams, num_beam_groups, and diversity_penalty.
• Compatibility: Works with decoder-only transformer models, matching the group_beam_search functionality found in transformers<4.56.0.
• Post-processing: Integrates DoLa contrastive scoring for post-processing logits before token selection, further refining output quality.

Good For

• Applications requiring creative text generation where diverse outputs are preferred.
• Scenarios where avoiding repetitive or generic responses is crucial.
• Exploring a wider array of potential continuations in sequence generation tasks.