💧 LiqFit

LiqFit is an easy-to-use framework for few-shot learning of cross-encoder models. These models are trained to determine whether two statements entail, contradict each other, or are neutral — a task setup applicable to various information extraction problems, such as text classification, named entity recognition, and question answering. With LiqFit, competitive results can be achieved using as few as 8 examples per label.

For detailed information and access to the code, visit the GitHub repository.

---

Key Features

• 🔢 Few-shot capable
  Achieve high performance with only 8 examples per label. Significantly improves default zero-shot classifiers.

• 📝 Task-agnostic inference setting
  Leverages natural language inference (NLI) to generalize across various information extraction tasks, including NER and QA.

• 🌈 Generalizes to unseen classes
  Models maintain the ability to classify labels not present in the training set, thanks to pre-finetuning on large NLI/classification corpora.

• ⚙️ Supports multiple cross-encoder types
  Compatible with standard, binary, and encoder-decoder cross-encoder architectures.

• ⚖️ Robust to class imbalance
  Normalization techniques ensure stability and performance even with unbalanced datasets.

• 🏷️ Multi-label classification support
  Designed for both multi-class and multi-label classification tasks.

---

Benchmarks:

Model & examples per label	Emotion	AgNews	SST5
Comprehend-it/0	56.60	79.82	37.9
Comprehend-it/8	63.38	85.9	46.67
Comprehend-it/64	80.7	88	47
SetFit/0	57.54	56.36	24.11
SetFit/8	56.81	64.93	33.61
SetFit/64	79.03	88	45.38

Limitations

• 🤔 Computational overhead
  Requires N forward passes through the transformer for N labels.