THETA-embeddings

Pre-computed embeddings generated by THETA, a domain-specific embedding model fine-tuned on Qwen3-Embedding for sociology and social science texts.

Description

This dataset contains dense vector embeddings produced under three settings:

• zero_shot: Embeddings from the base Qwen3-Embedding model without fine-tuning
• supervised: Embeddings from the LoRA-adapted model trained with label-guided contrastive learning
• unsupervised: Embeddings from the LoRA-adapted model trained with SimCSE

Repository Structure

CodeSoulco/THETA-embeddings/
├── 0.6B/
│   ├── zero_shot/
│   ├── supervised/
│   └── unsupervised/
└── 4B/
    ├── zero_shot/
    ├── supervised/
    └── unsupervised/

Embedding Details

Model	Dimension	Format
Qwen3-Embedding-0.6B	896	.npy
Qwen3-Embedding-4B	2560	.npy

Source Datasets: germanCoal, FCPB, socialTwitter, hatespeech, mental_health

How to Use

import numpy as np

# Load pre-computed embeddings
embeddings = np.load("0.6B/zero_shot/germanCoal_zero_shot_embeddings.npy")
print(embeddings.shape)  # (num_samples, 896)

Or download via huggingface_hub:

from huggingface_hub import hf_hub_download
import numpy as np

path = hf_hub_download(
    repo_id="CodeSoulco/THETA-embeddings",
    filename="0.6B/supervised/socialTwitter_supervised_embeddings.npy",
    repo_type="dataset"
)
embeddings = np.load(path)

Related

• Model (LoRA weights): CodeSoulco/THETA

License

This dataset is released under the MIT License.

Citation

@misc{theta2026,
  title={THETA: Textual Hybrid Embedding--based Topic Analysis},
  author={CodeSoul},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/CodeSoulco/THETA}
}