All You Need to Know about SAMSum Dataset

All You Need to Know about SAMSum Dataset


Are you a researcher or developer interested in the field of dialogue summarization? If so, you won’t want to miss the groundbreaking SAMSum Dataset— a unique dataset that is poised to transform the state of the art.

In this blog post, referencing the paper “SAMSum Corpus: A Human-annotated Dialogue Dataset for Abstractive Summarization”, we’ll take a deep dive into the SAMSum Dataset, uncovering its key features and exploring how you can leverage this powerful resource with your LLM API. Whether you’re looking to fine-tune language models, benchmark summarization approaches, or simply stay ahead of the curve, this comprehensive overview has you covered. Let’s dive in!

What Is SAMSum Dataset?


The SAMSum Corpus, or SAMSum Dataset, was created by researchers from the Samsung R&D Institute Poland — Bogdan Gliwa, Iwona Mochol, Maciej Biesek, and Aleksander Wawer.


The dialogues in the SAMSum Corpus are in English.

Data Structure

  • Data Instances: The dataset contains 16,369 chat dialogues. Here is an example dialogue and summary from the SAMSum Corpus:
  • Data Fields: Each dialogue instance includes the actual dialogue text, with each utterance labeled with the speaker’s name. Each dialogue also has a manually written abstractive summary.
  • Data Splits: The dataset is split into 14,732 dialogues for training, 818 for validation, and 819 for testing.

Source Data

Since there was no existing dataset of messenger-style conversations available, the researchers decided to create the SAMSum Dataset from scratch. Linguists fluent in English were asked to construct natural-sounding chat dialogues reflecting the topics and styles typical of real-world messenger conversations.

Data Annotators

The paper “SAMSum Corpus: A Human-annotated Dialogue Dataset for Abstractive Summarization” does not explicitly mention the identities of the data annotators for the SAMSum Dataset. It states that the dialogues were created by “linguists fluent in English” and that the manual summaries were also written by “language experts”.

So the data annotators were likely professional linguists and language experts recruited by the researchers at Samsung R&D Institute Poland to construct the dialogues and write the summaries. However, their specific identities are not provided in the paper.

Why Did People Create SAMSum Dataset?

The authors note that major research efforts in text summarization have so far focused on summarizing single-speaker documents like news articles, due to the availability of large, high-quality news datasets with summaries. However, a comprehensive dataset for dialogue summarization was lacking.

The authors argue that the challenges posed by abstractive dialogue summarization require dedicated models and evaluation approaches, beyond what has been developed for news summarization. By creating the SAMSum Corpus, the researchers aimed to provide a high-quality dataset of chat dialogues with manual abstractive summaries, which can be used by the research community to further study and advance dialogue summarization.

How Can I Finetune My LLM With SAMSum Dataset?

Here are the steps you can follow to fine-tune a large language model (LLM) using the SAMSum dataset:

Step 1: Obtain an LLM API

  • Sign up for an API key or access token to use the LLM in your code.
  • Novita AI offers developers a diverse range of LLM API options, providing access to cutting-edge models like llama-3–8b-instruct, llama-3–70b-instruct, mistral-7b-instruct, and hermes-2-pro-llama-3–8b. 
  • Additionally, adjustable parameters such as top-p, temperature, presence penalty, and max tokens enable you to customize the performance of the LLM.
  • You can freely compare and evaluate these different LLM choices on Novita AI Playground, helping you select the most suitable model for your specific needs.

Step 2: Download the SAMSum dataset

  • The SAMSum dataset is available for download on Hugging Face.
  • Follow the instructions to download the dataset and unpack the files.

Step 3: Preprocess the data

  • The SAMSum dataset contains dialogues and their corresponding abstractive summaries.
  • You’ll need to preprocess the data to be compatible with the input and output formats expected by your LLM.
  • This may involve tokenizing the text, separating the dialogues and summaries, and potentially adding special tokens or formatting.

Step 4: Fine-tune the LLM

  • Depending on the LLM you’re using, the fine-tuning process may differ slightly.
  • Generally, you’ll need to fine-tune the model on the SAMSum dataset, using the dialogues as the input and the summaries as the target output.
  • This can be done using the LLM’s fine-tuning API or by implementing a custom training loop.
  • You may need to experiment with different hyperparameters, such as learning rate, batch size, and number of training epochs, to achieve the best performance.

Step 5: Evaluate the fine-tuned model

  • Use the test set from the SAMSum dataset to evaluate the performance of your fine-tuned model.
  • Metrics like ROUGE scores, as used in the original paper, can be helpful for assessing the quality of the generated summaries.
  • You may also want to perform manual evaluation or human evaluation to get a better sense of the model’s performance.

Step 6: Iterate and improve

  • Based on the evaluation results, you may need to adjust your fine-tuning process, try different LLM architectures, or explore other techniques to improve the model’s performance on dialogue summarization.
  • The SAMSum dataset provides a valuable resource for iterating and advancing the state-of-the-art in this task.

What Are the Limitations of Samsum Dataset?

Based on the reasearch paper by Gliwa et al. (2019), here are some of the key limitations of the SAMSum dataset:

Limited Diversity of Dialogues

  • The dialogues in the SAMSum dataset were created by linguists, rather than being sourced from real-world chat conversations.
  • While the researchers aimed to make the dialogues reflect typical messenger conversations, the dataset may not capture the full breadth and diversity of real-world chat interactions.
  • The dialogues may lack the nuances and idiosyncrasies that naturally occur in spontaneous conversations.

Potential Bias in Summaries

  • The summaries for the dialogues were also written by language experts, rather than being sourced from real users.
  • This means the summaries may reflect the biases and perspectives of the annotators, rather than representing how actual users would summarize the conversations.
  • The summaries may also be influenced by the instructions given to the annotators, such as the requirement to include names of interlocutors and be written in the third person.

Limited Size

  • The SAMSum dataset, while relatively large compared to some other dialogue summarization datasets, is still relatively small compared to news summarization datasets like CNN/Daily Mail.
  • The limited size of the dataset may constrain the ability of models to learn robust and generalizable dialogue summarization capabilities.

Lack of Contextual Information

  • The dataset only includes the dialogue text and summary, without any additional contextual information about the participants, the topic of the conversation, or the setting.
  • This lack of contextual information may limit the ability of models to capture the nuances and implications of the dialogues.

Potential Noise and Inconsistencies

  • Despite the cleaning process, the dataset may still contain some noise, typos, or inconsistencies, as it was created manually by linguists.
  • This could introduce challenges for models trying to learn patterns and generalize from the data.

Overall, the SAMSum dataset represents a valuable contribution to the field of dialogue summarization research, but it also has some inherent limitations that researchers should be aware of when using and evaluating the dataset. Addressing these limitations may be an area for future work in expanding and enhancing dialogue summarization datasets.


The SAMSum Dataset represents an important contribution to the field of dialogue summarization research. By providing a high-quality dataset of messenger-style conversations with manual abstractive summaries, the creators have aimed to spur further advancements in this area.

However, the dataset also has some inherent limitations that researchers should be aware of, such as the synthetic nature of the dialogues, potential biases in the summaries, and the relatively small size compared to news summarization datasets.

Addressing these limitations and further expanding the dataset may be valuable areas for future work. Overall, the SAMSum Dataset is a valuable resource that can help drive progress in the challenging task of abstractive dialogue summarization.


Gliwa, B., Mochol, I., Biesek, M., & Wawer, A. (2019). SAMSum Corpus: A Human-annotated Dialogue Dataset for Abstractive Summarization. arXiv preprint arXiv:1911.12237.

Novita AI is the all-in-one cloud platform that empowers your AI ambitions. With seamlessly integrated APIs, serverless computing, and GPU acceleration, we provide the cost-effective tools you need to rapidly build and scale your AI-driven business. Eliminate infrastructure headaches and get started for free — Novita AI makes your AI dreams a reality.
Recommended Reading
Red Pajama LLM: Public Dataset Analysis Revealed