Languages: en
Multilinguality: monolingual
Size Categories: 10K<n<100K
Licenses: cc-by-4.0
Language Creators: found
Annotations Creators: expert-generated
Source Datasets: extended|bookcorpus

Dataset Card for LAMBADA

Dataset Summary

The LAMBADA evaluates the capabilities of computational models for text understanding by means of a word prediction task. LAMBADA is a collection of narrative passages sharing the characteristic that human subjects are able to guess their last word if they are exposed to the whole passage, but not if they only see the last sentence preceding the target word. To succeed on LAMBADA, computational models cannot simply rely on local context, but must be able to keep track of information in the broader discourse.

The LAMBADA dataset is extracted from BookCorpus and consists of 10'022 passages, divided into 4'869 development and 5'153 test passages. The training data for language models to be tested on LAMBADA include the full text of 2'662 novels (disjoint from those in dev+test), comprising 203 million words.

Supported Tasks and Leaderboards

Long range dependency evaluated as (last) word prediction


The text in the dataset is in English. The associated BCP-47 code is en.

Dataset Structure

Data Instances

A data point is a text sequence (passage) including the context, the target sentence (the last one) and the target word. For each passage in the dev and the test splits, the word to be guessed is the last one.

The training data include the full text of 2'662 novels (disjoint from those in dev+test), comprising more than 200M words. It consists of text from the same domain as the dev+test passages, but not filtered in any way.

Each training instance has a category field indicating which sub-category the book was extracted from. This field is not given for the dev and test splits.

An example looks like this:

{"category": "Mystery",
 "text": "bob could have been called in at this point , but he was n't miffed at his exclusion at all . he was relieved at not being brought into this initial discussion with central command . `` let 's go make some grub , '' said bob as he turned to danny . danny did n't keep his stoic expression , but with a look of irritation got up and left the room with bob",

Data Fields

  • category: the sub-category of books from which the book was extracted from. Only available for the training split.
  • text: the text (concatenation of context, target sentence and target word). The word to be guessed is the last one.

Data Splits

  • train: 2'662 novels
  • dev: 4'869 passages
  • test: 5'153 passages

Dataset Creation

Curation Rationale

The dataset aims at evaluating the ability of language models to hold long-term contextual memories. Instances are extracted from books because they display long-term dependencies. In particular, the data are curated such that the target words are easy to guess by human subjects when they can look at the whole passage they come from, but nearly impossible if only the last sentence is considered.

Source Data

Initial Data Collection and Normalization

The corpus was duplicated and potentially offensive material were filtered out with a stop word list.

Who are the source language producers?

The passages are extracted from novels from Book Corpus.


Annotation process

The authors required two consecutive subjects (paid crowdsourcers) to exactly match the missing word based on the whole passage (comprising the context and the target sentence), and made sure that no subject (out of ten) was able to provide it based on local context only, even when given 3 guesses.

Who are the annotators?

The text is self-annotated but was curated by asking (paid) crowdsourcers to guess the last word.

Personal and Sensitive Information

[More Information Needed]

Considerations for Using the Data

Social Impact of Dataset

[More Information Needed]

Discussion of Biases

[More Information Needed]

Other Known Limitations

[More Information Needed]

Additional Information

Dataset Curators

[More Information Needed]

Licensing Information

The dataset is released under the [CC BY 4.0](Creative Commons Attribution 4.0 International) license.

Citation Information

  author    = {Paperno, Denis  and  Kruszewski, Germ\'{a}n  and  Lazaridou,
Angeliki  and  Pham, Ngoc Quan  and  Bernardi, Raffaella  and  Pezzelle,
Sandro  and  Baroni, Marco  and  Boleda, Gemma  and  Fernandez, Raquel},
  title     = {The {LAMBADA} dataset: Word prediction requiring a broad
discourse context},
  booktitle = {Proceedings of the 54th Annual Meeting of the Association for
Computational Linguistics (Volume 1: Long Papers)},
  month     = {August},
  year      = {2016},
  address   = {Berlin, Germany},
  publisher = {Association for Computational Linguistics},
  pages     = {1525--1534},
  url       = {}


Thanks to @VictorSanh for adding this dataset.

Update on GitHub
Explore dataset Edit dataset tags

Models trained or fine-tuned on lambada

None yet