credence-distilbert-uncased-20251223-234956-hatexplain

CREDENCE checkpoint for hatexplain (toxicity) with backbone distilbert-base-uncased (5 concept heads).

Paper: https://huggingface.co/papers/2604.24170 · arXiv: 2604.24170

Training run folder: distilbert-uncased_20251223_234956

Files

• credence_checkpoint.pt — PyTorch checkpoint (model_state_dict, config, metadata, optimizer state)

Load with huggingface_hub

from huggingface_hub import hf_hub_download
import torch

path = hf_hub_download(repo_id="tankiit/credence-distilbert-uncased-20251223-234956-hatexplain", filename="credence_checkpoint.pt")
ckpt = torch.load(path, map_location="cpu", weights_only=False)
state = ckpt["model_state_dict"]
config = ckpt["config"]

Load into your CREDENCE model implementation (see project credence.py).

Citation

@article{mukherjee2026credence,
  title={Credal Concept Bottleneck Models for Epistemic--Aleatoric Uncertainty Decomposition},
  author={Mukherjee, et al.},
  journal={arXiv preprint arXiv:2604.24170},
  year={2026}
}

Evaluation (test)

Metric	Value
Accuracy	0.5826
ρ(epistemic, error)	0.1044
ρ(aleatoric, unknown)	0.2455