Open-source R and J++ neuron lenses for Qwen3.5-4B ship with runnable readout tooling
burny_tech · x · 2026-10-29
bubskee released Apache-2.0 R and J++ workspace lenses for the Qwen3.5-4B text decoder on Hugging Face, with FP32 transport matrices per layer, a standalone reader, and frozen eval results. J++ reportedly improves concept recovery on heldout tasks, with multilingual gains mostly from relations and input concepts; readouts are framed as hypotheses, not causal neuron labels. The reader supports CUDA/MPS/CPU, and the repo ships install plus readlenses.py commands that print logit, R, and J++ top-10 readouts for any prompt, with options to pick a lens or rank the full vocabulary.
More from Research
- Falcon: lightweight streaming audio-visual speaker extraction with RealSSA benchmark — thu-ml · 2026-10-29
- Dev post-trains Qwen3-4B into a tool-using agent: 66 MB llama.cpp LoRA runs on a Mac, val 0.51→0.83 — DesignerIsland5524 · 2026-10-29
- Zero-HLM trains humanoid loco-manipulation with zero real robot data, lifting success 23% to 69% — siyuanhuang95 · 2026-10-29
- Matt Clancy floats $1bn AI project to check the robustness of all published research — sebkrier · 2026-10-29
- The heat equation is Gaussian blur: how 19th-century math underpins computer vision — techNmak · 2026-10-29
- Anthropic partners with six drug developers and launches its own drug discovery efforts with Claude — elonmusk · 2026-10-29