Fewclue Eval

fewclue-eval

FewCLUE: A Chinese Few-shot Learning Evaluation Benchmark — Xu et al. (2021) (arXiv:2107.07498, 2021)

What this evaluates

Evaluates Chinese NLP models on few-shot learning across nine tasks, including single-sentence classification, sentence-pair classification, and machine reading comprehension. It tests the ability of pre-trained language models and few-shot prompting/fine-tuning methods to generalize with limited labeled data.

Datasets

• FewCLUE — total ?; splits: train (-1), test (-1); repo https://github.com/CLUEbenchmark/FewCLUE

Metrics

• accuracy (primary) — range: [0, 1]
  • Proportion of correctly predicted labels out of total instances. Calculated as correct predictions divided by total test instances.

Input / output format

Input: Chinese text sentences or sentence pairs, optionally reformatted into cloze-style prompts with [MASK] tokens and candidate label words for few-shot methods.

Output: Predicted label token or probability distribution over candidate label tokens.

Scoring recipe

correct = 0
for pred, gold in zip(predictions, gold_labels):
    if pred == gold:
        correct += 1
return correct / len(gold_labels)

Common pitfalls

• Template and label-word selection significantly impact few-shot performance in PET/ADAPET.
• Small labeled subsets cause high variance; results should be averaged over multiple random seeds/splits.
• Cloze-style reformulation may alter semantic meaning, affecting model predictions.

Evidence (verbatim from paper)

Finally, supervised training is performed on the dataset by using cross-entropy loss.

Citation

@misc{xu2021fewclue,
  title={FewCLUE: A Chinese Few-shot Learning Evaluation Benchmark},
  author={Xu et al. (2021)},
  year={2021},
  note={arXiv:2107.07498}
}

• arXiv: 2107.07498