annoying-ai-2b

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

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "your-username/annoying-ai-2b"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

prompt = "User: Can you help me write an email?\nAI:"
inputs = tokenizer(prompt, return_tensors="pt")
output = model.generate(**inputs, max_new_tokens=80)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Training Details

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.