GCUL: Guided Clustering-based Uncertain Learning -- released models

Checkpoints of the GCUL project. One repository, several sub-directories -- see the table below. The code, the processed splits and the figures live in the GitHub repository (https://github.com/u678868/GCUL).

Sub-models

path	files	task / dataset	notes
distilbert/kaggle/	config.json, model.safetensors, tokenizer.json, tokenizer_config.json	11-class emotion classification (Kaggle)	DistilBertForSequenceClassification, test acc 89.37 %, macro-F1 0.8958
distilbert/goemotions/	same	28-class (GoEmotions)	test acc 39.03 %, macro-F1 0.3037
distilbert/dair/	same	6-class (dair-ai/emotion)	test acc 92.90 %, macro-F1 0.9042
bilstm/	config.json, model.safetensors, word2idx.json	11-class (Kaggle)	BiLSTM + self-attention over GloVe-300d
textcnn/	config.json, model.safetensors, word2idx.json	11-class (Kaggle)	TextCNN, 25k vocabulary
gcul_pca/	config.json, phase1.safetensors, phase3.safetensors	11-class + uncertain (Kaggle)	PCA-10 clustering
gcul_lda/	config.json, phase1.safetensors, phase3.safetensors	11-class + uncertain (Kaggle)	LDA-10 clustering

phase1 is the K-class model; phase3 is the retrained (K+1)-class model that can emit the uncertain label (id = 11).

Usage

from transformers import AutoModelForSequenceClassification, AutoTokenizer

repo = "u678868/GCUL"
model = AutoModelForSequenceClassification.from_pretrained(f"{repo}", subfolder="distilbert/kaggle")
tok = AutoTokenizer.from_pretrained(f"{repo}", subfolder="distilbert/kaggle")
print(model.config.id2label)          # 0 -> "anger", ... (from data/splits/label_mapping.csv)

For the non-transformer sub-models, load the tensors with safetensors.torch.load_file and instantiate the matching class from src/models/ of the GitHub repository (config.json carries the hyper-parameters).

Vocabularies

word2idx.json maps token strings to integer ids with string keys; convert on load:

import json
word2idx = {int(k): v for k, v in json.load(open("word2idx.json")).items()}

The vocabulary is released as JSON rather than pickle on purpose (no pickle-version issues, and it is human-inspectable).

Label encoding

Class ids follow the per-dataset mapping in data/splits/label_mapping.csv of the GitHub repository (single source of truth). They are not uniformly alphabetical: Kaggle is pure alphabetical, GoEmotions is alphabetical with neutral forced to the last index, and dair-ai uses the original integer identity.

Conversion notes

• Tensor names are preserved from the original .pt checkpoints, with one exception: transformers

  = 5 writes LayerNorm.{gamma,beta} where the source files used LayerNorm.{weight,bias}. Both spellings load via from_pretrained; tensor values are bit-identical.

• distilbert-base-uncased is not mirrored here -- it is a public checkpoint.
• The sklearn LabelEncoder stored inside the original TextCNN checkpoint was discarded: its classes_ are identical to the released class_names in config.json.
• _convert.py is a development-side tool that converts the original .pt checkpoints to .safetensors. It expects the full project tree (with saved_models/ and datasets/) to be present as siblings. End users only need the released .safetensors files and do not need to run this script.

Citation

@misc{fan2026gcul,
  title  = {GCUL: Guided Clustering-based Uncertain Learning for Multi-Class Emotion Classification},
  author = {Fan, Zhongqi and Zhang, Tianyou and Chen, Fei},
  year   = {2026},
  note   = {https://github.com/u678868/GCUL}
}