Qwen3.5 2B Bumped

--- library_name: transformers license: apache-2.0 license_link: https://huggingface.co/Qwen/Qwen3.5-2B/blob/main/LICENSE pipeline_tag: image-text-to-text base_model: - Qwen/Qwen3.5-2B countDownloads: - config.json --- # Qwen3.5-2B-Bumped *What happens when you physically "bump" a neural network?* <img width="400px" src="https://cdn-uploads.huggingface.co/production/uploads/632c02c725a261e762ce7ab6/kY0D-9ycBskm2u87HG3da.jpeg"> ### The Experiment Imagine a neural network's weights as a delicate grid of liquid-filled vials meticulously arranged on a pristine laboratory table. Every drop of liquid represents the precision of a floating-point number. Now, imagine someone accidentally bumps the table. The core structure remains standing, but the delicate surface levels slosh and spill. This model is an experiment in light to extreme weight perturbation and structural robustness. In technical terms, the Least Significant Bits of this model's weights have been entirely overwritten with essentially random data. <strike>The critical top bits are always preserved ensuring the model mathematically survives the "spill" without triggering `NaN` or `Inf` calculation crashes. </strike> Moved to a custom llama.cpp fork so I don't have to worry about this. ### Performance & Behavior **This is an experimental proof-of-concept, not a daily driver.** Because the lowest 1 to 15 bits of the weights have been completely replaced by noise, the model operates similar to a quantization. * Expect inference quality roughly equivalent to an aggressive `q4` or `q5` quantization at best. It works, but the "spilled" precision means it will be more prone to hallucinations and such. The models are based on Qwen3.5-2B-Base so they need to be finetuned to use for chatting.

Source context: 107 downloads · 0 likes · Pipeline image-text-to-text · Library transformers · Repo yideda/Qwen3.5-2B-Bumped