SODA2207/MafiaMD

Mafia-MD

Multimodal group discussions, public events, and subsequent human actions

Mafia-MD is a research resource built from naturally occurring, Russian-language online Mafia games. It links speaker-side audio and video to structured public game events, making it possible to study how information from a group discussion relates to a participant's later recorded action.

Overview of the corpus and research tasks. Faces are synthetic illustrations; identifying example fields have been omitted. The diagram depicts the broader corpus and proposed tasks, not only the files in the current feature release.

Corpus	Scale
Games	2,937
Public-speech clips	93,882
Recorded discussion	approximately 1,438 hours
Eligible decisions in the paper's primary benchmark	26,235

Release status: This repository currently contains feature embeddings and clip metadata, not the complete forecasting benchmark. Source audio/video, transcript text, decision-level labels, public-event inputs and fixed task splits are not included in the present package. The complete benchmark, including eligible source media and task materials, is planned for release by camera-ready, subject to the applicable permissions. The separate 400-game replication cohort is not included in the current feature package.

Research question

Can representations of an ongoing group discussion improve predictions of an individual's subsequent action beyond what is already known from public events?

In the paper's primary task, a model observes the permitted discussion and public game state before voting and predicts whether a participant's first vote differs from their earlier nomination. Nominations and votes are recorded game actions, rather than annotations of a person's intentions.

The label is a nomination–vote mismatch. It does not establish deception, persuasion, a change of belief, or any other hidden psychological state. The task is specific to online Mafia; transfer to other settings has not been established. Decision labels and benchmark splits are not yet included in this repository.

Illustrative speaker-side example: face-blurred video, an audio waveform, an English translation of an ASR excerpt and selected public nominations. The waveform and excerpt are not time-aligned. Source media and transcripts are not part of the current feature release.

What the paper reports

Model input	Historical AUROC	Game-disjoint replication AUROC
Public events (TabM)	0.721	0.709
+ Target participant's transcript (T)	0.749	0.735
+ Expanded discussion branch (O)	0.764	0.753

The O branch combines other participants' transcripts with expanded public context. In matched input comparisons, the speech-specific increment is positive on the replication cohort, while the historical confidence interval includes zero. These are predictive comparisons, not causal effects. The replication cohort is retrospective and comes from the same platform; game-disjoint splits do not guarantee unseen participants.

The paper also audits visual representations. The tested representations reflect capture conditions and respond to some speech-related visual signals, but their tested additions do not consistently improve the main forecast. This result is limited to the evaluated models, target and speaker-side recordings; it does not establish that useful visual information is absent.

What's in this release

compact/                    Clip-aligned NumPy feature arrays
  ib_audio.npy              ImageBind audio embeddings
  ib_video.npy              ImageBind video embeddings
  emo_audio.npy             Pooled GigaAM-Emo audio features
  emo_video.npy             Pooled EmotiEff visual features
  e5.npy                    Text embeddings, partial query-prefix coverage
  e5_noprefix.npy           Text embeddings, no prefix
  e5_query_available.npy    Availability mask for e5.npy
  e5_noprefix_available.npy Availability mask for e5_noprefix.npy
metadata/
  clips.jsonl               Clip manifest and row alignment
  release.json              Artifact inventory, checksums and feature recipes
  text_coverage.json        Text-feature coverage information
raw/shards/                 Tar shards with original feature arrays

The raw/shards/ directory contains feature arrays, not source media. ImageBind arrays are float32 vectors; compact emotion features concatenate the frame mean and population standard deviation as float32. Compact text arrays are float16. The unpooled emotion arrays in the shards retain frame-level information. See metadata/release.json for exact recipes and checksums.

Each line in metadata/clips.jsonl describes one clip. Its row field aligns with the first dimension of the compact arrays. Game IDs, clip IDs and stems are pseudonymous; the manifest does not contain transcript text, participant account identifiers or original local paths.

Text-feature coverage

e5.npy contains query-prefix features only for a subset of the clips. Exclude unavailable rows using e5_query_available.npy. The no-prefix features in e5_noprefix.npy also have an associated e5_noprefix_available.npy mask. Do not treat missing rows or zero-filled rows as valid embeddings.

Quick start

Download and inspect an audio embedding array and the clip manifest:

import json
import numpy as np
from huggingface_hub import hf_hub_download

repo = "SODA2207/MafiaMD"
audio_path = hf_hub_download(repo, "compact/ib_audio.npy", repo_type="dataset")
manifest_path = hf_hub_download(repo, "metadata/clips.jsonl", repo_type="dataset")

audio = np.load(audio_path, mmap_mode="r", allow_pickle=False)
with open(manifest_path, encoding="utf-8") as f:
    clip = json.loads(next(f))

print("Array shape:", audio.shape)
print("Clip metadata:", clip)
print("Aligned audio embedding:", audio[clip["row"]])

For text features, load and apply the matching availability mask before analysis. Validate downloaded artifacts against the SHA256 checksums in the release manifest.

Scope and responsible use

• Recordings capture the active speaker, not simultaneous listener reactions. Recording quality and capture conditions vary.
• The corpus comes from a single platform and is Russian-language. Results should not be assumed to transfer to other populations, languages or interaction settings.
• The paper's pre-vote boundary relies in part on platform rules and event-log semantics; manual validation of event alignment and ASR remains incomplete.
• Public-speech clips include ordinary speeches and eligible re-speeches. Forecasting requires task-specific filtering at the observation cutoff; the feature package alone does not supply all inputs needed to reproduce the benchmark.
• Pseudonymization does not guarantee anonymity. Audio and visual embeddings can retain identity-related information. Do not attempt to identify participants or link representations to real identities. Access and use remain subject to the dataset's applicable access conditions; this feature release grants no additional media redistribution rights.

Paper and citation

Mafia-MD: Forecasting Human Decisions from Multimodal Group Discussion. The paper describes the benchmark protocol, baseline models, evaluation cohorts and limitations. Citation details will be added when the paper's publication information is available.