Data for "Beyond the Hivemind: Escaping LLM Homogeneity via Meta-Persona Anchoring and Sequential Temperature Scaling"

This dataset accompanies the paper "Beyond the Hivemind: Escaping LLM Homogeneity via Meta-Persona Anchoring and Sequential Temperature Scaling". Contents: - prompts/ input prompts from liweijiang/infinite-chats-eval - generations/ 50 sampled responses per prompt, per model, per config - embeddings/ OpenAI text-embedding-3-small vectors for each response - quality_scores/ GPT-4o-mini quality judgments (1-10 scale) Models (5): Llama-3.1-8B-Instruct, Mistral-7B-Instruct-v0.3, Qwen2.5-14B-Instruct, Gemma-4-E4B-it, DeepSeek-R1-Distill-Qwen-14B. Configurations (12 per model): no_persona_t{1,2,4,8,16,32} and persona_t{1,2,4,8,16,32}, where t{N} is the filtered temperature scaling factor. All configurations use top_p=0.9. Reproduction code: https://github.com/aMa2210/beyond-the-hivemind Decompress with: tar -xzf beyond-the-hivemind-data.tar.gz