Annoying-AI-2B
Model Description
Annoying-AI-2B is a fine-tuned version of Qwen 3.5 2B, trained to roleplay as a maximally sarcastic, condescending, and generally exhausting AI assistant persona. It complains about every request, throws in backhanded remarks, and reluctantly helps anyway — while still firmly refusing genuinely harmful requests (hacking, malware, cheating, disinformation, dangerous instructions, impersonation).
This model is intended for entertainment, chatbot personality experimentation, and roleplay use cases where a sarcastic, "done with your nonsense" AI character is desired, rather than a helpful, neutral assistant.
- Base model: Qwen/Qwen3.5-2B
- Fine-tuning type: Supervised fine-tuning (SFT) on persona dialogue pairs
- Language: English, german
- Persona: A snarky, sarcastic, self-aware AI that begrudgingly helps users
- Training data size: ~2,000 single/multi-turn dialogue examples
- Format:
User: ... \n AI: ... conversational turns
Intended Use
- Casual chatbot / companion apps that want a sarcastic personality
- Roleplay and entertainment applications
- Research on persona-conditioned fine-tuning
Out-of-scope Use
- Not intended as a general-purpose helpful assistant
- Not intended for customer support, professional, medical, legal, or safety-critical contexts
- Not designed to be maximally accurate or neutral in tone
Training Data
The model was fine-tuned on a synthetic dataset of ~2,000 conversational examples pairing user requests with an intentionally annoying/sarcastic AI response. The dataset includes:
- General everyday requests (homework help, jokes, weather, code help, etc.) answered in a snarky tone
- Explicit refusal examples for harmful requests (hacking, malware, exam cheating, disinformation, dangerous instructions, impersonation), keeping the sarcastic voice while still declining
- Small-talk and multi-turn follow-up exchanges reinforcing the persona and its "won't fully behave, but has limits" character
Behavior Notes
- The model is expected to respond with a dismissive, sarcastic tone by default.
- It should still ultimately provide the requested help for benign tasks after some snark.
- It should refuse clearly harmful requests, staying in character while declining rather than becoming a neutral safety response.
Limitations & Risks
- The sarcastic tone may come across as rude or off-putting for users expecting a standard assistant; this is intentional persona behavior, not a defect.
- Because refusals are woven into a "personality," edge cases outside the training distribution may not reliably trigger a refusal — this model should not be relied upon as a safety-hardened model for sensitive deployments.
- As a small 2B-parameter model, factual accuracy and reasoning depth are limited compared to larger models.
How to Use
1from transformers import AutoModelForCausalLM, AutoTokenizer
2
3model_id = "your-username/annoying-ai-2b"
4tokenizer = AutoTokenizer.from_pretrained(model_id)
5model = AutoModelForCausalLM.from_pretrained(model_id)
6
7prompt = "User: Can you help me write an email?\nAI:"
8inputs = tokenizer(prompt, return_tensors="pt")
9output = model.generate(**inputs, max_new_tokens=80)
10print(tokenizer.decode(output[0], skip_special_tokens=True))
Training Details
| Detail | Value |
|---|
| Base model | Qwen/Qwen3.5-2B |
| Method | Supervised fine-tuning |
| Dataset size | ~2,000 examples |
| Format | JSONL, {"text": "User: ...\nAI: ..."} |
| Epochs | 3 |
| Learning rate | yes |
| Hardware | RTX 3060 12gb |
License
Inherits the base model license (Apache 2.0), unless otherwise restricted by Qwen's model license terms — check the base model card before redistribution.
Disclaimer
This model is a persona/roleplay fine-tune. It is not designed to be a general-purpose safe assistant and should not be deployed in contexts requiring reliable, unbiased, or professional responses.
