README.md

---
license: other
license_name: mixed-per-scenario
license_link: https://huggingface.co/datasets/DuplexGen/duplexgen-corpus/blob/main/README.md#heritage--licensing
pretty_name: DuplexGen Corpus
task_categories:
  - text-generation
  - conversational
language:
  - en
tags:
  - turn-taking
  - dialogue
  - backchannel
  - full-duplex
  - spoken-dialogue
configs:
  - config_name: INT_dialogues
    data_files:
      - split: train
        path: dialogues/INT/train.jsonl
  - config_name: INT_annotations
    data_files:
      - split: train
        path: annotations/INT/train.jsonl
      - split: test
        path: annotations/INT/test.jsonl
  - config_name: NEG_dialogues
    data_files:
      - split: train
        path: dialogues/NEG/train.jsonl
  - config_name: NEG_annotations
    data_files:
      - split: train
        path: annotations/NEG/train.jsonl
      - split: test
        path: annotations/NEG/test.jsonl
  - config_name: PER_dialogues
    data_files:
      - split: train
        path: dialogues/PER/train.jsonl
  - config_name: PER_annotations
    data_files:
      - split: train
        path: annotations/PER/train.jsonl
      - split: test
        path: annotations/PER/test.jsonl
  - config_name: PLN_dialogues
    data_files:
      - split: train
        path: dialogues/PLN/train.jsonl
  - config_name: PLN_annotations
    data_files:
      - split: train
        path: annotations/PLN/train.jsonl
      - split: test
        path: annotations/PLN/test.jsonl
  - config_name: SOC_dialogues
    data_files:
      - split: train
        path: dialogues/SOC/train.jsonl
  - config_name: SOC_annotations
    data_files:
      - split: train
        path: annotations/SOC/train.jsonl
      - split: test
        path: annotations/SOC/test.jsonl
  - config_name: TEA_dialogues
    data_files:
      - split: train
        path: dialogues/TEA/train.jsonl
  - config_name: TEA_annotations
    data_files:
      - split: train
        path: annotations/TEA/train.jsonl
      - split: test
        path: annotations/TEA/test.jsonl
---

# DuplexGen Corpus

Text corpus for **DuplexGen: Adaptive Synthesis of Human–AI Turn-Taking
Dialogues**.

This dataset contains DuplexGen-generated dialogues and our own human
turn-taking slot annotations, used to train and calibrate models that
predict when a listener should take the floor, backchannel, or stay silent
during spoken conversation.

A companion dataset, **[`DuplexGen/duplexgen-spoken`](https://huggingface.co/datasets/DuplexGen/duplexgen-spoken)**,
provides a spoken-audio rendering of the generated dialogues (via
Chatterbox TTS). The generation pipeline and training/eval code are at
[github.com/duplexgen/duplexgen-code](https://github.com/duplexgen/duplexgen-code).

## Links

- 📄 **Paper** — arXiv link coming soon
- 💻 **Code** — [github.com/duplexgen/duplexgen-code](https://github.com/duplexgen/duplexgen-code)
- 🎛️ **Finetune** — [github.com/duplexgen/personaplex-finetune](https://github.com/duplexgen/personaplex-finetune)
- 🤗 **Corpus** — [DuplexGen/duplexgen-corpus](https://huggingface.co/datasets/DuplexGen/duplexgen-corpus)
- 🤗 **Spoken** — [DuplexGen/duplexgen-spoken](https://huggingface.co/datasets/DuplexGen/duplexgen-spoken)
- 🌐 **Demo** — [duplexgen.github.io](https://duplexgen.github.io/)

## Dataset summary

Two parts, laid out per scenario code (`TEA`, `PLN`, `INT`, `NEG`, `PER`,
`SOC`):

```
dialogues/<CODE>/train.jsonl        # DuplexGen-generated dialogues (train-only)
annotations/<CODE>/train.jsonl      # human slot-level turn-taking annotations
annotations/<CODE>/test.jsonl
```

**Only DuplexGen-generated dialogues and our own human slot annotations are
released here — no third-party raw source text is redistributed.** See
[Heritage / licensing](#heritage--licensing) below for how each scenario's
prompts/situations were seeded.

## Loading

This repo defines **one config per scenario × part**: a `<CODE>_dialogues`
config (single `train` split) and a `<CODE>_annotations` config (`train` +
`test` splits) for each scenario code `{TEA, PLN, INT, NEG, PER, SOC}`:

```python
from datasets import load_dataset

# generated dialogues for one scenario (train-only)
ds = load_dataset("DuplexGen/duplexgen-corpus", "PER_dialogues", split="train")

# human annotations for the same scenario, test split
ann = load_dataset("DuplexGen/duplexgen-corpus", "INT_annotations", split="test")
```

Valid configs:
- `{TEA,PLN,INT,NEG,PER,SOC}_dialogues` — split `train` only
- `{TEA,PLN,INT,NEG,PER,SOC}_annotations` — splits `train`, `test`

## Data fields

### `dialogues/<CODE>/train.jsonl`

Each line is one generated dialogue:

```
{
  "example_id": str,
  "scenario": str,          # scenario code, e.g. "TEA"
  "license": str,           # inherited upstream license, see Heritage / licensing below
  "context": str,
  "style": str,             # e.g. "spoken"
  "disfluency_target": str, # which speaker role disfluency was targeted at
  "speakers": [str, ...],
  "history": [
    {
      "role": str,
      "content": str,
      "segments": [
        # a segment is EITHER a plain content span:
        {"full_content": str},
        # OR a per-word turn-taking decision slot:
        {
          "word_index": int,
          "probs": {
            "floor_taking": float,
            "backchannel": float,
            "silence": float
          },
          "decision": str,          # the sampled/selected action at this slot
          "inserted_token": str | null
        }
      ]
    },
    ...
  ]
}
```

`segments` interleaves plain-text spans with per-word turn-taking decision
slots in document order — replaying a turn's `segments` list reconstructs
`content` plus the inserted turn-taking behavior at each slot.

### `annotations/<CODE>/{train,test}.jsonl`

Each line is one dialogue's **human** slot-level turn-taking annotation.
Note the deliberate terminology difference from `dialogues`: annotations
use `"silent"`/`"take_floor"` where dialogues use `"silence"`/`"floor_taking"`
(`"backchannel"` is shared).

```
{
  "example_id": str,
  "scenario": str,
  "license": str,
  "history": [
    {
      "role": str,
      "content": str,
      "boundaries": [
        {
          "word_index": int,
          "total_count": int,        # number of human raters for this slot
          "counts": {                # raw vote counts, subset of:
            "silent": int,
            "backchannel": int,
            "take_floor": int
          },
          "probabilities": {         # counts normalized by total_count, over:
            "silent": float,
            "backchannel": float,
            "take_floor": float
          }
        },
        ...
      ]
    },
    ...
  ]
}
```

## Dataset statistics (verified)

**`dialogues` (train-only, generated):**

| Scenario | Count |
|---|---|
| TEA | 1000 |
| PLN | 1000 |
| INT | 1000 |
| NEG | 1000 |
| PER | 999 |
| SOC | 1000 |
| **Total** | **5999** |

Total per-word turn-taking decision segments (slots with `word_index`/`probs`,
excluding plain `full_content` segments) across all 5999 dialogues:
**125,137**.

**`annotations` (human-labeled):**

| Split | Per scenario | Total (× 6 scenarios) |
|---|---|---|
| train | 20 | 120 |
| test | 50 | 300 |
| **Total** | **70** | **420** |

## Heritage / licensing

Each scenario's dialogues were seeded from a different upstream source and
therefore carries a different license. **Only the DuplexGen-generated
dialogue text and our own human annotations are distributed in this
dataset — no raw text from these upstream sources is redistributed here.**

| Scenario | Code | Source dataset | License |
|---|---|---|---|
| Socratic teaching | TEA | SocraticLM | Apache-2.0 |
| Planning | PLN | MultiWOZ | MIT |
| Interview | INT | Anthropic Interviewer | CC-BY-4.0 |
| Negotiation | NEG | CraigslistBargain | MIT |
| Persuasion | PER | DailyPersuasion | Apache-2.0 |
| Social chat | SOC | SODA | CC-BY-4.0 |

**Attribution (PER):** the `PER` scenario is seeded from DailyPersuasion
(PersuGPT), released under Apache-2.0. Please retain attribution to
DailyPersuasion when redistributing the `PER` split.

Because licensing differs per scenario, this dataset's top-level Hub
`license` metadata is set to `other`; the table above is the authoritative
per-scenario license reference. If you use only a subset of scenarios,
comply with that scenario's license only.

## Generator credit

- Dialogue **text** (spoken-style conversion and synthesized turn-taking
  dialogues) is generated by **Qwen3.5-122B-A10B**.
- Spoken **audio** (in the companion [`DuplexGen/duplexgen-spoken`](https://huggingface.co/datasets/DuplexGen/duplexgen-spoken)
  dataset) is rendered via **Chatterbox** TTS.
- **No third-party raw text** is included in this dataset — only
  DuplexGen-generated dialogues and our own human slot annotations are
  released.

## Citation

If you use this dataset, please cite the DuplexGen paper.