Nitral-Archive
/

Qwen3.5-4B_Homebrew

Model card Files Files and versions

Trained on a local 3090:

1: 4bit qlora to loosen both censorship/refusals, and thinking tags use. (1 epoch at 7e-5, [sharegpt chatml] using 4chan/reddit and toxic dpo data).

2: Abliterated using a custom fork of heretic.

Downloads last month: 73

Safetensors

Model size

5B params

Tensor type

BF16

·

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Nitral-Archive/Qwen3.5-4B_Homebrew

Quantizations