Qwen3.5-2B-abliterated-Fable5-Vibecoding

This model is a specialized fine-tune of huihui-ai/Huihui-Qwen3.5-2B-abliterated designed to act as an uncensored, agentic coding assistant and software architect.

Because the base model is "abliterated" (refusal vectors surgically removed), this model will provide direct, unfiltered, and highly technical responses without moralizing or refusing complex coding requests.

🧠 Training Data & Methodology

The model was trained using Supervised Fine-Tuning (SFT) via LoRA on a combined dataset of ~1GB of high-quality coding traces:

1. Fable Traces Cleaned: Contains complex agentic workflows, multi-step reasoning, bash command executions, file reading, and tool-calling syntax.
2. Vibe-Coding-Claude-Fable-5: Contains high-level software architecture scenarios, debugging strategies (e.g., infinite React re-renders), and practical implementation plans.

💡 Recommended System Prompt

For the best results and to trigger the agentic/architect persona, use the following system prompt:

You are an expert, uncensored AI coding assistant and software architect. You provide direct, practical implementation plans and use tools when necessary.

🚀 How to Use Using Transformers (Python)

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "Toaster496/Qwen3.5-2B-abliterated-Fable5-Vibecoding"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

messages = [
    {"role": "system", "content": "You are an expert, uncensored AI coding assistant and software architect. You provide direct, practical implementation plans and use tools when necessary."},
    {"role": "user", "content": "React app re-renders infinitely. Scenario 2: Provide a practical implementation plan."}
]

inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to(model.device)

outputs = model.generate(inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Using llama.cpp / Ollama (Local Inference) Note: To run this locally on consumer hardware, convert the Safetensors to GGUF format using llama.cpp or Unsloth.

# Example using llama-cli
llama-cli -hf Toaster496/Qwen3.5-2B-abliterated-Fable5-Vibecoding-GGUF -cnv

⚠️ Disclaimer This model is "abliterated" and fine-tuned on uncensored, agentic datasets. It may produce content that is unfiltered, highly opinionated, or attempt to execute system commands without safety guardrails. Always use this model in sandboxed environments when testing agentic tool-calling capabilities or executing generated bash scripts. 🙏 Acknowledgements Base Model: huihui-ai for the abliterated Qwen3.5 weights. Datasets: thejagstudio and attentionAllYouNeed for the exceptional coding and agentic trace datasets.