ada-flo/nlp-hack-debate

---
language:
- en
- ko
license: cc-by-4.0
task_categories:
- text2text-generation
pretty_name: Debate-Themed Dialogue Generation Dataset
size_categories:
- 10K<n<100K
---

# ada-flo/nlp-hack-debate

Bilingual (English + Korean) training data for an LSTM-based seq2seq debate
chatbot. Each record is a `(topic, input_context, target_output)` triple
plus precomputed `encoder_input` / `decoder_input` / `decoder_target` ready
for seq2seq training.

## Schema

```json
{
"id": "ibm_argq_30k_8b4b12caccad",
"lang": "en",
"source": "ibm_argq_30k",
"is_synthetic": false,
"input_stance": "pro",
"target_stance": "con",
"topic": "We should abandon marriage",
"input_context": "abandoning marriage allows for people to grow as themselves...",
"target_output": "committment and stability are important in the lives of children...",
"encoder_input": "We should abandon marriage <SEP> abandoning marriage allows...",
"decoder_input": "<SOS> committment and stability are important...",
"decoder_target": "committment and stability are important... <EOS>",
"meta": { "source_record_ids": [], "quality_input_WA": 1.0, "...": "..." }
}
```

Top-level fields filterable in the HF dataset viewer:

| Field | Values |
|---|---|
| `lang` | `en`, `ko` |
| `source` | `ibm_argq_30k`, `mc_conversation`, `isotonic_conversation`, `casual_conversation`, `ko_debate_synth`, `korean_petitions` |
| `is_synthetic` | `true`, `false` |
| `input_stance` | `pro`, `con`, `petition_position`, `supportive`, `oppositional`, … |
| `target_stance` | `pro`, `con`, `opposition`, … |

## Splits

| Split | Records | EN | KO |
|---|---|---|---|
| train | 40,006 | 27,093 | 12,913 |
| validation | 5,050 | 3,811 | 1,239 |
| test | 4,429 | 2,760 | 1,669 |

Splits are **topic-level** for debate-shaped sources (motion-grouped records
all land in one split — no leakage). Casual chat and topic-seeded synth use
row-wise split because they share placeholder topics.

## Sources (train split)

| Source | Records |
|---|---|
| ibm_argq_30k | 24,126 |
| korean_petitions | 8,203 |
| ko_debate_synth | 4,710 |
| isotonic_conversation | 1,186 |
| mc_conversation | 971 |
| casual_conversation | 810 |

## Source descriptions

- **ibm_argq_30k** — [IBM Argument Quality Ranking 30K](https://huggingface.co/datasets/ibm-research/argument_quality_ranking_30k). Real human pro/con stance pairs over ~70 motions.
- **mc_conversation** — [mc-ai/conversation_dataset](https://huggingface.co/datasets/mc-ai/conversation_dataset), filtered to `corpus_id=persuasionforgood`. Real persuasion-themed multi-turn dialogue (Persuasion-for-Good corpus).
- **isotonic_conversation** — [Isotonic/human_assistant_conversation](https://huggingface.co/datasets/Isotonic/human_assistant_conversation), filtered to single-turn rows without dialog markers or code-task content.
- **casual_conversation** — [SohamGhadge/casual-conversation](https://huggingface.co/datasets/SohamGhadge/casual-conversation). Casual greeting-style exchanges for conversational fluency.
- **ko_debate_synth** — Topic-seeded debate-pair synthesis (Korean). 98 curated debate motions × 30 LLM-generated PRO/CON pairs each. Uses Qwen3-235B-A22B-Instruct via vLLM at temperature 0.9. Both directions per pair.
- **korean_petitions** — Korean Petitions corpus (청와대 국민청원 2017–2019, via Korpora). Petition title = motion, body (truncated to 280 chars) = `input_context`, vLLM-synthesized counter-argument = `target_output`.

## Synthetic data

Records with `meta.is_synthetic=true` were generated by Qwen3-235B-A22B-Instruct
served via vLLM. Synthesis prompt versions are recorded in
`meta.synthesis_prompt_version`.

| Prompt version | Used by |
|---|---|
| v1 (counterargument) | korean_petitions |
| v1 (debate_pair) | ko_debate_synth |

Prompts: see `src/synth/prompts.py` in the source repository.

## License

CC BY 4.0. Source corpora retain their original licenses; consult each source
link above for redistribution terms before commercial use.

## Repository

Generated by https://github.com/ada-flo/nlp-hack — see that repo for the
full preprocessing pipeline, source adapters, and synth client code.