Enhanced Helpster - With Embeddings

This dataset extends the HelpSteer dataset (35,331 prompts) with computed embeddings and entropy metrics derived from the Llama 2 70B language model. For each prompt, we extracted high-dimensional embeddings (8192 dimensions) capturing semantic representations, and calculated entropy measures during text generation to quantify model uncertainty. The entropy metrics include: average entropy of the first 10 generated tokens, minimum and maximum entropy values, confidence measures (complement of entropy), and first token probability. These features were computed using generation parameters (temperature=1, top_k=1) to access the full probability distribution over the vocabulary. The dataset preserves all original HelpSteer annotations (helpfulness, correctness, coherence, complexity, verbosity on 0-4 scales) while adding computational features that enable research on active learning strategies, uncertainty-based case selection, and the relationship between model uncertainty and human evaluation metrics. This enhanced dataset facilitates studies on efficient annotation sampling, entropy-based selection methods, and semantic clustering approaches for large language model evaluation.