fodorad's picture
Upload README.md with huggingface_hub
3aac7cd verified
|
Raw
History Blame Contribute Delete
4.67 kB
metadata
license: other
license_name: chalearn-fi-derived
license_link: https://chalearnlap.cvc.uab.cat/dataset/24/description/
pretty_name: PersonalityLinMulT  FI Extracted Features (fi.h5)
tags:
  - personality
  - big-five
  - multimodal
  - affective-computing
  - feature-extraction
task_categories:
  - audio-classification
  - video-classification
size_categories:
  - 10K<n<100K

PersonalityLinMulT — First Impressions V2 extracted features (fi.h5)

Pre-extracted multimodal features for the First Impressions V2 (FI) dataset, packaged as a single HDF5 file for training the models in PersonalityLinMulT: Big Five personality perception and interview-variable (hireability) estimation.

⚠️ These are derived features (embeddings), not the source videos. No raw ChaLearn First Impressions video, audio, or frames are re-shared here. This file contains only numerical feature vectors extracted from the official dataset. To obtain the raw videos, register for and download the dataset from the official ChaLearn LAP release under its own license terms.

What this is for

Training reads this file and nothing else — it is the reproducibility boundary for the project. Pull it and train immediately, with no need to obtain the raw videos or re-run the multi-hour GPU feature extraction:

# In a clone of https://github.com/fodorad/PersonalityLinMulT
make pull-fi-h5        # downloads fi.h5 into data/processed/FI/h5/
make train-fi-bigfive  # train Big Five regression

Or download directly:

from huggingface_hub import hf_hub_download
path = hf_hub_download(
    "fodorad/personalitylinmult-fi", "fi.h5", repo_type="dataset"
)

Contents

fi.h5 — 10 000 clips (6000 train / 2000 valid / 2000 test), ~33 GB, float16. Row-indexed and feature-major: __getitem__(i) reads exactly one chunk per feature. Every sequence feature carries a mask (True = the protagonist was detected / a real token) and a length.

feature modality source model shape (N, T, D)
wavlm acoustic WavLM Base+ (layer 9) (10000, 750, 768)
emotion2vec acoustic emotion2vec+ (10000, 750, 768)
dinov2 visual DINOv2 Base (full frame) (10000, 375, 768)
emotieffnet visual EmotiEffNet (EfficientNet-B0) (10000, 375, 1280)
farl visual FaRL (ViT-B/16) (10000, 375, 512)
xlm_roberta_gt textual XLM-RoBERTa tokens, ground-truth (10000, 128, 768)
mmbert_gt textual mmBERT tokens, ground-truth (10000, 128, 768)

Visual features (emotieffnet, farl) track the protagonist's face; the mask marks frames where the protagonist was detected. dinov2 is dense (full frame). Text is token-level (T, 768) (not pooled) for both XLM-RoBERTa and mmBERT — the representation a cross-modal sequence model attends over — padded to 128 tokens with the mask marking real tokens. Both are extracted over the ground-truth transcript.

Labels stored per clip: the Big Five traits (openness, conscientiousness, extraversion, agreeableness, neuroticism, and emotional_stability = 1 − neuroticism) plus the interview variable, all in [0, 1].

Provenance and reproducibility

The file's HDF5 root attributes embed the builder config verbatim, the git commit SHA of the code that produced it (with a -dirty suffix if built from an uncommitted tree), and the source label CSV's SHA-256 — so any copy traces back to the exact recipe. To rebuild from scratch and verify:

make pipeline-fi   # decode → extract features → build fi.h5 (requires raw FI)

Citation

Features derived from the ChaLearn First Impressions V2 dataset. If you use this artifact, please cite the original dataset and the PersonalityLinMulT paper:

Fodor et al., Multimodal Sentiment and Personality Perception Under Speech: A Comparison of Transformer-based Architectures. PMLR v173

Links