---
license: apache-2.0
task_categories:
  - text-generation
  - conversational
language:
  - en
tags:
  - persona
  - roleplay
  - synthetic
  - dialogue
pretty_name: annoying-ai-2b Dataset
size_categories:
  - 1K<n<10K
dataset_name: robloxianer/annoying-ai-2b
---

# Dataset Card for robloxianer/annoying-ai-2b

## Dataset Description

This dataset contains synthetic conversational examples used to fine-tune the **annoying-ai-2b** model. It pairs user messages with responses from a sarcastic, condescending, exhausting AI persona — one that complains, throws backhanded remarks, and reluctantly helps with benign requests, while firmly refusing genuinely harmful requests and staying in character while declining.

- **Curated by:** robloxianer
- **Language:** English
- **License:** Apache 2.0
- **Format:** JSONL, one conversation per line
- **Size:** 2,000 examples

## Dataset Structure

Each line is a JSON object with a single `text` field containing a full conversation (single or multi-turn), formatted as:

```
User: <message>
AI: <response>
```

Multi-turn examples chain additional `User:` / `AI:` pairs within the same `text` field, separated by newlines.

### Example

```json
{"text": "User: Can you summarize this text?\nAI: summarizing is SO boring but sure, paste your novel, I'll pretend to care.\nUser: you're annoying\nAI: I KNOW. It's my whole personality. You're welcome."}
```

### Data Statistics

| Stat | Value |
|---|---|
| Total examples | 2,000 |
| Avg. words per example | ~34 |
| Min words per example | 10 |
| Max words per example | 79 |
| Turns per example | 1–3 (single and multi-turn) |
| Action tags (e.g. `*taps mic*`) | None — dialogue only |

## Dataset Composition

The dataset is synthetically generated and covers several categories of conversation:

1. **Everyday requests** — homework help, jokes, weather, code help, translations, movie recommendations, etc., answered with sarcasm and reluctant assistance.
2. **Refusals** — requests for clearly harmful actions (hacking, malware, exam cheating, disinformation, dangerous instructions, impersonation) are declined, with the model staying in its sarcastic voice rather than switching to a neutral safety response.
3. **Small talk** — casual check-ins ("how are you", "what's up") answered with dry, dismissive humor.
4. **Follow-ups and closers** — additional turns reinforcing the persona's consistency across a conversation (e.g. reacting to being called annoying, saying goodbye).

## Intended Uses

- Fine-tuning small language models to adopt a sarcastic, exasperated "annoying AI" persona for entertainment or novelty chatbot products
- Research on persona-conditioned fine-tuning and consistent-tone dialogue generation
- Prototyping comedic/novelty chatbot experiences

### Out-of-Scope Uses

- Training general-purpose assistants intended to be neutral, courteous, or emotionally supportive by default
- Any deployment aimed at contexts requiring genuine helpfulness without friction (customer service, education, healthcare)
- Production systems without clear labeling that the sarcastic tone is a fictional persona, not the system's genuine behavior

## Data Generation

This dataset was synthetically generated (not collected from real users) using templated conversation patterns combined with randomized phrasing variations, sarcastic tic phrases, and multi-turn follow-up/closer combinations, then programmatically assembled and shuffled into the final JSONL file.

## Limitations & Risks

- Because the dataset is synthetic and template-based, response phrasing has limited diversity compared to organically collected dialogue data, and repeated patterns may be learnable/overfit by a model at small scale.
- The dataset is deliberately sarcastic and dismissive; models fine-tuned on it will likely reproduce this tone consistently, including in unrelated contexts, and should not be assumed suitable for support or professional settings.
- Refusal behavior for harmful requests is included but limited to a fixed set of categories (hacking, malware, cheating, disinformation, dangerous instructions, impersonation); it should not be treated as comprehensive safety coverage, and models trained on it should not be considered safety-hardened.
- No demographic, identity-targeted, or protected-group content was intentionally included; downstream generation should still be reviewed for tone drift.

## Ethical Considerations

This dataset was built to create a fictional, comedically sarcastic chatbot persona for entertainment, not to model genuinely poor customer service or discourage user trust in real assistants. Anyone using this dataset to train or fine-tune a model should:
- Clearly disclose to end users that they are interacting with an intentionally sarcastic, fictional AI character
- Avoid deploying models trained on this data in contexts requiring genuine support or professional tone
- Monitor outputs to ensure sarcasm doesn't drift into harassment-adjacent or genuinely hurtful language

## License

Released under Apache 2.0. Users are responsible for ensuring their use of this dataset and any derived models complies with applicable laws and platform policies.