---
license: mit
language:
- en
task_categories:
- summarization
tags:
- dialogue-summarization
- dialogsum
- samsum
size_categories:
- 10K<n<100K
configs:
- config_name: default
  data_files:
  - split: train
    path: train.jsonl
  - split: validation
    path: validation.jsonl
  - split: test
    path: test.jsonl
---

# Dialogue Summarization Dataset

Short conversation → abstractive summary pairs, used to fine-tune
[`tuanhqv123/longt5-meeting-summarization`](https://huggingface.co/tuanhqv123/longt5-meeting-summarization).

A cleaned merge of two public dialogue-summarization benchmarks: **DialogSum** and **SAMSum**.

## Fields

| Field    | Type   | Description                              |
|----------|--------|------------------------------------------|
| `input`  | string | The dialogue / conversation              |
| `output` | string | Reference abstractive summary            |
| `source` | string | `dialogsum` or `samsum`                  |

## Splits

Each source's **original** train/val/test split is preserved (no re-shuffling across splits, so
the splits stay comparable to published DialogSum/SAMSum results). The combined ratio is **~88 / 4 / 8**:

| Split      | Total  | %     | dialogsum | samsum |
|------------|--------|-------|-----------|--------|
| train      | 26,095 | 88.1% | 12,453    | 13,642 |
| validation | 1,256  | 4.2%  | 500       | 756    |
| test       | 2,259  | 7.6%  | 1,489     | 770    |

Conversations are short: input median ≈ 73–116 words, p95 ≈ 400 tokens — no long context needed.

## Cleaning

Built from the raw datasets with a discovery-driven pipeline (see the
[training repo](https://huggingface.co/tuanhqv123/longt5-meeting-summarization/tree/main/code)):

- normalize unicode quotes/dashes, strip emojis (SAMSum chat noise)
- drop auto-detected disfluencies (`uh, um, hmm…`)
- **unify speaker tags:** DialogSum's `#Person1#` → `Person1`, matching SAMSum's bare-name style
- filter: input ≥ 20 words, output 5–200 words, drop exact duplicates

Kept ~96% of rows (30,828 → 29,610).

## Sources & licenses

Both are public benchmark datasets (no proprietary/internal data):

| `source` tag | train rows | origin | license |
|--------------|-----------|--------|---------|
| `dialogsum`  | 12,453 | [DialogSum](https://github.com/cylnlp/dialogsum) — two-person dialogues with `#PersonN#` tags | MIT |
| `samsum`     | 13,642 | [SAMSum](https://huggingface.co/datasets/knkarthick/samsum) — messenger-style group chats | CC-BY-NC-ND 4.0 |

Please respect the original licenses. Cite the original works when using this data.

## Usage

```python
from datasets import load_dataset

ds = load_dataset("tuanhqv123/meeting-summarization")
print(ds)
print(ds["train"][0]["input"][:300])
print(ds["train"][0]["output"])
```