Sign in to view source links and access this dataset
Description
Each JSON file contains one full game of Buzzwords & Misdemeanors, a simulated courtroom hearing. The dataset records the Game Master's raw structured decisions per beat, including speaker, beat type, fact_index clue channel, intensity, and stage. Authored by BastienHot, the dataset was last updated on June 14, 2026.
Use Cases
Analyze agent decision sequences based on the structured beat-by-beat trace data.
Study the interaction between deterministic guards and language model actors based on the described courtroom sequencing invariants.
Train or evaluate agent models on structured, multi-turn dialogue tasks using the recorded game traces.
Investigate the effect of different actor styles (per-style LoRA) on dialogue generation within a constrained game environment.
Strengths
Traces capture structured decisions per beat, including multiple specific fields like speaker and intensity.
Data generation employs deterministic guards to enforce sequencing invariants, suggesting consistency.
Each record represents a complete game session, providing context for multi-turn interactions.
Limitations
Column-level documentation is absent; field semantics must be inferred after download.
Row count is unknown, which may limit suitability assessment.
The description metadata is limited; actual data quality requires manual inspection after download.
Provenance
Source
huggingface
Collection Method
Generated by a Game Master (MiniCPM5-1B + a distilled director LoRA, GBNF-constrained) directing a hearing over a sampled truth, with actors delivering lines.
Freshness
Last updated 2026-06-14 14:23:51; freshness should be verified.
License is unknown; terms of use must be verified before application.