Usage tips & FAQ
#1
by paijo77 - opened
FAQ & Usage Tips
Q: How do I run this with Ollama?
ollama run hf.co/paijo77/qwen3-4b-abliterated
Q: Does thinking mode still work?
Yes! The <think></think> tokens are preserved. Abliteration only removes the refusal direction, not the reasoning capability.
Q: What's the difference vs the base Qwen3-4B?
This model cannot refuse requests. The refusal behavior is removed at the weight level — no system prompt can re-enable it. KL divergence of 0.0388 means minimal impact on general capability.
Q: Can I fine-tune this further?
Yes, use our dataset: https://huggingface.co/datasets/paijo77/berkahkarya-id-finance-dataset
Q: How was this made?
Using Heretic on a GTX 1660 SUPER (6GB VRAM). Full writeup on r/LocalLLaMA.
Support the creator: 👉 https://www.tip.md/oyi77