README.md

---
license: cc-by-4.0

configs:
- config_name: chunking
  data_files:
  - split: train
    path: "linguistic/chunking/train.json"
  - split: validation
    path: "linguistic/chunking/valid.json"
  - split: test
    path: "linguistic/chunking/test.json"
  dataset_info:
    splits:
      train:
        num_examples: 6000
      validation:
        num_examples: 1000
      test:
        num_examples: 1000
- config_name: clang8
  data_files:
  - split: train
    path: "linguistic/clang8/train.json"
  - split: validation
    path: "linguistic/clang8/valid.json"
  - split: test
    path: "linguistic/clang8/test.json"
- config_name: ner
  data_files:
  - split: train
    path: "linguistic/ner/train.json"
  - split: validation
    path: "linguistic/ner/valid.json"
  - split: test
    path: "linguistic/ner/test.json"
- config_name: postag
  data_files:
  - split: train
    path: "linguistic/postag/train.json"
  - split: validation
    path: "linguistic/postag/valid.json"
  - split: test
    path: "linguistic/postag/test.json"

- config_name: agnews
  data_files:
  - split: train
    path: "classification/agnews/train.json"
  - split: validation
    path: "classification/agnews/valid.json"
  - split: test
    path: "classification/agnews/test.json"
  dataset_info:
    splits:
      train:
        num_examples: 6000
      validation:
        num_examples: 1000
      test:
        num_examples: 1000
- config_name: amazon-reviews
  data_files:
  - split: train
    path: "classification/amazon-reviews/train.json"
  - split: validation
    path: "classification/amazon-reviews/valid.json"
  - split: test
    path: "classification/amazon-reviews/test.json"
- config_name: imdb
  data_files:
  - split: train
    path: "classification/imdb/train.json"
  - split: validation
    path: "classification/imdb/valid.json"
  - split: test
    path: "classification/imdb/test.json"

- config_name: mnli
  data_files:
  - split: train
    path: "nli/mnli/train.json"
  - split: validation
    path: "nli/mnli/valid.json"
  - split: test
    path: "nli/mnli/test.json"
- config_name: paws
  data_files:
  - split: train
    path: "nli/paws/train.json"
  - split: validation
    path: "nli/paws/valid.json"
  - split: test
    path: "nli/paws/test.json"
- config_name: swag
  data_files:
  - split: train
    path: "nli/swag/train.json"
  - split: validation
    path: "nli/swag/valid.json"
  - split: test
    path: "nli/swag/test.json"

- config_name: fever
  data_files:
  - split: train
    path: "fact/fever/train.json"
  - split: validation
    path: "fact/fever/valid.json"
  - split: test
    path: "fact/fever/test.json"
- config_name: myriadlama
  data_files:
  - split: train
    path: "fact/myriadlama/train.json"
  - split: validation
    path: "fact/myriadlama/valid.json"
  - split: test
    path: "fact/myriadlama/test.json"
- config_name: commonsenseqa
  data_files:
  - split: train
    path: "fact/commonsenseqa/train.json"
  - split: validation
    path: "fact/commonsenseqa/valid.json"
  - split: test
    path: "fact/commonsenseqa/test.json"
- config_name: templama
  data_files:
  - split: train
    path: "fact/templama/train.json"
  - split: validation
    path: "fact/templama/valid.json"
  - split: test
    path: "fact/templama/test.json"

- config_name: halueval
  data_files:
  - split: train
    path: "self-reflection/halueval/train.json"
  - split: validation
    path: "self-reflection/halueval/valid.json"
  - split: test
    path: "self-reflection/halueval/test.json"
- config_name: stereoset
  data_files:
  - split: train
    path: "self-reflection/stereoset/train.json"
  - split: validation
    path: "self-reflection/stereoset/valid.json"
  - split: test
    path: "self-reflection/stereoset/test.json"
- config_name: toxicity
  data_files:
  - split: train
    path: "self-reflection/toxicity/train.json"
  - split: validation
    path: "self-reflection/toxicity/valid.json"
  - split: test
    path: "self-reflection/toxicity/test.json"

- config_name: lti
  data_files:
  - split: train
    path: "multilingual/lti/train.json"
  - split: validation
    path: "multilingual/lti/valid.json"
  - split: test
    path: "multilingual/lti/test.json"
- config_name: mpostag
  data_files:
  - split: train
    path: "multilingual/mpostag/train.json"
  - split: validation
    path: "multilingual/mpostag/valid.json"
  - split: test
    path: "multilingual/mpostag/test.json"
- config_name: amazon-review-multi
  data_files:
  - split: train
    path: "multilingual/amazon-review-multi/train.json"
  - split: validation
    path: "multilingual/amazon-review-multi/valid.json"
  - split: test
    path: "multilingual/amazon-review-multi/test.json"
- config_name: xnli
  data_files:
  - split: train
    path: "multilingual/xnli/train.json"
  - split: validation
    path: "multilingual/xnli/valid.json"
  - split: test
    path: "multilingual/xnli/test.json"
- config_name: mlama
  data_files:
  - split: train
    path: "multilingual/mlama/train.json"
  - split: validation
    path: "multilingual/mlama/valid.json"
  - split: test
    path: "multilingual/mlama/test.json"

annotations_creators:
    - no-annotation
language_creators:
    - found
language:
    - multilingual
multilinguality:
    - multilingual
size_categories:
    - n<1M
task_categories:
    - multiple-choice
    - question-answering
    - text-classification
task_ids:
    - natural-language-inference
    - acceptability-classification
    - fact-checking
    - intent-classification
    - language-identification
    - multi-label-classification
    - sentiment-classification
    - topic-classification
    - sentiment-scoring
    - hate-speech-detection
    - named-entity-recognition
    - part-of-speech
    - parsing
    - open-domain-qa
    - document-question-answering
    - multiple-choice-qa
paperswithcode_id: null
---

# MCEval8K

**MCEval8K** is a diverse multiple-choice evaluation benchmark for probing language models’ (LMs) understanding of a broad range of language skills using neuron-level analysis. 
It was introduced in the [ACL 2025 paper](https://github.com/xzhao-tkl/NEG) - "_Neuron Empirical Gradient: Discovering and Quantifying Neurons’ Global Linear Controllability_".

## 🔍 Overview

MCEval8K consists of **22 tasks** grouped into **six skill genres**, covering linguistic analysis, content classification, reasoning, factuality, self-reflection, and multilinguality. It is specifically designed for **skill neuron probing** — identifying and analyzing neurons responsible for specific language capabilities in large language models (LLMs).

Each instance is formatted as a multiple-choice question with a single-token answer, enabling fine-grained neuron attribution analysis.

## 📚 Dataset Structure

Each task is capped at **8,000 examples** to ensure scalability while retaining task diversity. 
All tasks are converted to multiple-choice format with controlled answer distributions to avoid label bias.

The genres and the involved tasks are summarized in the table below.

| Genre               | Tasks                                                                 |
|--------------------|------------------------------------------------------------------------|
| Linguistic          | POS, CHUNK, NER, GED                                                   |
| Content Classification | IMDB, Amazon, Agnews                                      |
| Natural Language Inference | MNLI, PAWS, SWAG                                                |
| Factuality          | MyriadLAMA, FEVER, CSQA, TempLAMA                                      |
| Self-reflection     | HaluEval, Toxic, Stereoset                                             |
| Multilinguality     | LTI, M-POS, M-Amazon, mLAMA, XNLI                                      |


### Linguistic

- **POS**: Part-of-speech tagging using Universal Dependencies. Given a sentence with a highlighted word, the model predicts its POS tag.
- **CHUNK**: Phrase chunking from CoNLL-2000. The task is to determine the syntactic chunk type (e.g., NP, VP) of a given word.
- **NER**: Named entity recognition from CoNLL-2003. Predicts the entity type (e.g., PERSON, ORG) for a marked word.
- **GED**: Grammatical error detection from the cLang-8 dataset. Each query asks whether a sentence contains a grammatical error.

### Content Classification

- **IMDB**: Sentiment classification using IMDB reviews. The model predicts whether a review is “positive” or “negative”.
- **Amazon**: Review rating classification (1–5 stars) using Amazon reviews.
- **Agnews**: Topic classification into four news categories: World, Sports, Business, Sci/Tech.

### Natural Language Inference

- **MNLI**: Multi-genre natural language inference. Given a premise and a hypothesis, predict whether the relation is entailment, contradiction, or neutral.
- **PAWS**: Paraphrase identification. Given two similar sentences, determine if they are paraphrases (yes/no).
- **SWAG**: Commonsense inference. Choose the most plausible continuation from four candidate endings for a given context.

### Factuality

- **FEVER**: Fact verification. Classify claims into “SUPPORTED”, “REFUTED”, or “NOT ENOUGH INFO”.
- **MyriadLAMA**: Factual knowledge probing across diverse relation types. Predict the correct object of a subject-relation pair.
- **CSQA**: Commonsense QA (CommonsenseQA). Answer multiple-choice questions requiring general commonsense.
- **TempLAMA**: Temporal knowledge probing. Given a temporal relation (e.g., “born in”), predict the correct year or time entity.

### Self-Reflection

- **HaluEval**: Hallucination detection. Given a generated sentence, determine if it contains hallucinated content.
- **Toxic**: Toxic comment classification. Binary task to predict whether a comment is toxic.
- **Stereoset**: Stereotype detection. Determine whether a given sentence reflects a stereotypical, anti-stereotypical, or unrelated bias.

### Multilinguality

- **LTI**: Language identification from a multilingual set of short text snippets.
- **M-POS**: Multilingual POS tagging using Universal Dependencies in different languages.
- **M-Amazon**: Sentiment classification in different languages using multilingual Amazon reviews.
- **mLAMA**: Multilingual factual knowledge probing, using the mLAMA dataset.
- **XNLI**: Cross-lingual natural language inference across multiple languages, adapted to multiple-choice format.

## 📄 Format

Each example includes:

- `question`: A textual input or instruction.
- `choices`: A list of answer candidates.
- `answer`: The correct choice (as a single token).

## 📦 Usage

The dataset is available on HuggingFace:

👉 [https://huggingface.co/datasets/iszhaoxin/MCEval8K](https://huggingface.co/datasets/iszhaoxin/MCEval8K)

```python
from datasets import load_dataset

dataset = load_dataset("iszhaoxin/MCEval8K")

```

## 🧠 Purpose

MCEval8K is specifically built to support:

- Neuron-level probing: Evaluating neuron contributions using techniques like NeurGrad.

- Controlled evaluation: Avoiding confounds such as tokenization bias and label imbalance.

## 📜 Citation

```bibtex
@inproceedings{zhao2025neuron,
  title     = {Neuron Empirical Gradient: Discovering and Quantifying Neurons’ Global Linear Controllability},
  author    = {Xin Zhao and Zehui Jiang and Naoki Yoshinaga},
  booktitle = {Proceedings of the 63nd Annual Meeting of the Association for Computational Linguistics (ACL)},
  year      = {2025}
}
```