This dataset matters because sales interactions are sequential decisions: every message can change conversion probability. By providing turn-by-turn conversion trajectories, engagement/effectiveness scores and dense embeddings, this collection lets researchers and engineers analyze how specific utterances shift success likelihood and train agents that react in real time.
What Sets It Apart
- Turn-by-turn probability trajectories: tracks conversion probability at each message, so you can study causal effects of individual turns rather than just conversation-level labels.
- Rich annotations plus embeddings: includes outcome, engagement, sales-effectiveness scores and 3072-dim embeddings, so models can combine symbolic metrics with dense semantic features.
- Designed for RL use cases: synthetic scenarios and probability trajectories are structured to support reinforcement-learning agents that optimize mid-conversation actions and policies.
- Single train split, large scale: ~100k synthetic conversations concentrated on B2B SaaS, enabling large-scale model training without exposing real customer data.
Who It's For and Trade-offs
Great fit if you want to prototype or pretrain conversion-prediction models, analyze which message types most affect conversions, or train RL agents that require per-turn rewards. Look elsewhere if you need real customer data, multilingual coverage, or non-SaaS verticals: the data is synthetic, English-only, focused on B2B SaaS, and provided as a single train set requiring users to create validation/test splits.
