KSBCode/werewolf-bluffers

Ultimate Werewolf Bluffing Structured Datasets (SD1 & SD2)

Dataset Summary

This repository contains two structured datasets, Structured Dataset 1 (SD1) and Structured Dataset 2 (SD2), designed for the study of bluffing behavior in Ultimate Werewolf, a social deduction game. The datasets provide player-level behavioral representations extracted from game transcripts shared by slhleosun and bolinlai.

Each record represents a single player during a single game round, summarizing the player's behavior throughout that round rather than individual utterances.

Dataset Creation

Source Data

The structured datasets were generated from a temporary dataset whose utterances were previously annotated for bluffing behavior using DeepSeek API. Each sentence was classified according to whether it contained:

• Role Bluff
• Vote Bluff
• Both Role and Vote Bluff
• No Bluff

After bluff identification, the original 13 attributes were transformed into 35 behavioral features through feature engineering. These features include:

• Behavioral metrics

  • Number of accusations
  • Number of denials
  • Other conversational behavior statistics

• Bluff-related metrics

  • Number of bluffs
  • Bluff temporal distribution
  • Bluff frequency

• Interaction features

  • Voting targets
  • Player interaction statistics
  • Conversation dynamics

• Role-based features

  • Role changes
  • Werewolf indicator
  • Other role-related information

The data were then grouped by:

• Game
• Round
• Player

A player-round was considered a bluffer whenever at least one utterance during that round was classified as a bluff.

---

Structured Dataset 1 (SD1)

SD1 contains 3,782 player-round records, each represented by 35 engineered attributes.

Bluff Distribution

Bluff Type	Records
Role Bluff	2,603
Vote Bluff	973
Both Role and Vote Bluff	206

The imbalance between bluff types is expected because vote bluffs are generally easier for other players to verify during gameplay, making them less frequent than role bluffs.

Role Distribution

Role Category	Records
Werewolf	1,957
Other Roles	1,825

---

Statistical Analysis

The dataset was inspected through graphical exploration and statistical testing.

Although the Shapiro–Wilk test rejected global normality (statistic = 0.261, p-value = 5.124 × 10⁻¹⁷⁰), inspection of the model residuals showed that most attributes presented approximately normal or only slightly skewed residual distributions. This supports the application of Analysis of Variance (ANOVA) for feature relevance assessment.

Feature Selection

ANOVA Type II was employed to evaluate the statistical relevance of each engineered attribute.

Type II ANOVA was selected because:

• bluff categories are naturally unbalanced;
• the remaining variables present comparatively balanced distributions;
• Type II provides appropriate significance estimates under these conditions.

Attributes with:

• p-value > 0.05
• low F-statistic

were considered statistically non-significant and removed.

The discarded attributes were:

• Round duration
• Mean utterance length
• Number of indecisive statements
• Number of decisive statements
• Players mentioned
• Vote consistency rate

---

Structured Dataset 2 (SD2)

SD2 is a filtered version of SD1 containing the same 3,782 player-round records after removing statistically non-significant attributes identified through ANOVA Type II.

The objective of SD2 is to provide a cleaner feature space for machine learning algorithms while preserving all player-round observations.

Compared to SD1, SD2:

• contains fewer features;
• removes redundant or statistically irrelevant variables;
• maintains identical player and bluff distributions.

---

Potential Applications

These datasets can be used for:

• Deception detection
• Bluff classification
• Player profiling
• Behavioral pattern discovery
• Unsupervised clustering
• Explainable AI
• Social interaction modeling
• Human behavior analysis
• Conversational analytics
• Representation learning

---

Limitations

• Only player-rounds containing at least one identified bluff are included.
• The datasets originate from Ultimate Werewolf gameplay and may not generalize to other social deduction games.
• Bluff labels depend on the annotation methodology used during dataset construction.
• Bluff categories are naturally imbalanced, particularly between role and vote bluffs.

---

License

This dataset is distributed under the Apache License 2.0.

You are free to:

• use
• modify
• redistribute
• build upon

the dataset in accordance with the terms of the Apache 2.0 License.

For the full license text, see:

https://www.apache.org/licenses/LICENSE-2.0