---
license: other
task_categories:
  - text-generation
tags:
  - roleplay
  - persona
  - chat
  - character
  - tiny-slm
---

# roleplay-slm-mix

A curated multi-source conversation corpus for training small (sub-1B) language
models to do persona-consistent roleplay chat. All conversations are normalized to
a strict `User:` / `Assistant:` turn format with an optional persona/scene header.

Built for the TinyRole program (byte-level tiny SLMs).

## Files in this release (v1)

| File | Source | License | Convs |
|---|---|---|---|
| `data/coser-clean.txt` | [Neph0s/CoSER](https://huggingface.co/datasets/Neph0s/CoSER) | MIT | 76,004 |
| `data/mmrole-chat.txt` | [YanqiDai/MMRole](https://huggingface.co/datasets/YanqiDai/MMRole_dataset) | MIT | 3,900 |
| `data/character-llm-chat.txt` | [OpenMOSS/character-llm-data](https://huggingface.co/datasets/OpenMOSS-Team/character-llm-data) | **CC-BY-NC-4.0** | 12,929 |

**`character-llm-chat.txt` is NON-COMMERCIAL (CC-BY-NC-4.0).** It is a separate
file so the MIT-licensed files stay permissive. Do not use it commercially.

**Coming in v2:** RoleBench train split (Apache-2.0, ~85 MB — eval test split stays
out), PIPPA-filtered (Apache-2.0, pending NSFW filter pass).

**Deliberately excluded (unlicensed — cannot redistribute):** WorryFree scenarios,
hieunguyenminh/roleplay, AlekseyKorshuk synthetic trio. Used in training but not
redistributable.

## Format

Each conversation is a text block, blocks separated by a blank line:

```
[optional persona/scene header line]
User: <first user turn>
Assistant: <first assistant turn>
User: ...
Assistant: ...
```

Persona headers (when present) are a single capitalized line describing the
character/scene, NOT prefixed with a role.

## Per-source cleaning

- **coser-clean**: 771 classic-book characters. Strict User/Assistant alternation
  (merged consecutive same-role turns), name-prefix and stage-direction/asterisk
  stripping, dropped < 4-turn conversations. 305k raw → 76,004 clean.
- **mmrole-chat**: 85 characters, text-only extraction from the multimodal dataset
  (image references dropped; scene-description user turns retained). MIT.
- **character-llm-chat**: 9 historical/literary figures (Beethoven, Cleopatra,
  Caesar, Hermione, Newton, Socrates, Spartacus, Voldemort, MLK) via Experience
  Reconstruction (EMNLP 2023). Character role → Assistant, interlocutor → User.
  CC-BY-NC-4.0.

## Intended use

Fine-tuning small chat/roleplay models, as the final chat/persona stage of a
curriculum (grounding → narrative → this mix).

## Citation

Cite the original sources (above). Aggregation: TinyRole program, 2026.