Karsten Kuhnke PRO
mindchain
AI & ML interests
Mechanistic Interpretability, Sparse Autoencoders, JumpReLU, Reward Modeling, RLHF, AI Alignment, Function Calling, Gemma, Nemotron
Recent Activity
updated
a collection
about 12 hours ago
Text to Motion
liked
a model
about 12 hours ago
tencent/HY-Motion-1.0
updated
a Space
1 day ago
mindchain/react-blog
Organizations
Mamba/Transformers Combo Hybride
Google Gemma Scope 2: JumpReLU SAEs for Gemma 2 interpretability. 270M PT/IT, 1B PT variants. Neuronpedia integration. Mechanistic analysis.
Quantized Gemma 3 models: QAT for efficient deployment. Gemma-3-27B-IT Q4. Low memory, fast inference. Edge & production-ready LLMs.
NVIDIA Nemotron reward models: 340B, 8B BRRM, 70B/32B principle-based. RLHF training, preference learning, AI alignment research.
T5Gemma-2 encoder-decoder models: 270M, 1B, 4B sizes. Text-to-text, summarization, translation. Google's architecture for structured generation.
Hybrid Mamba + Transformer architectures. NVIDIA Nemotron-3-Nano-30B-A3B (32B). BF16 & GGUF. Efficient long-context & in-context learning.
-
nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16
Text Generation • 32B • Updated • 229k • 509 -
bartowski/nvidia_Nemotron-3-Nano-30B-A3B-GGUF
Text Generation • 32B • Updated • 9.16k • 8 -
nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16
Text Generation • 32B • Updated • 7.15k • 82 -
nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8
Text Generation • 32B • Updated • 440k • 214
Mamba/Transformers Combo Hybride
Top LLMs 2025: ZAI GLM-4.7 (358B) & Moonshot Kimi-K2-Thinking. Next-gen reasoning, code, multilingual. State-of-the-art performance. Production-ready.
Custom JumpReLU Sparse Autoencoders for mechanistic interpretability. T5Gemma-2 SAEs across all layers. AI safety & interpretability research.
NVIDIA Nemotron RL datasets for AI agent training. Web search, workplace tasks, instruction following, structured outputs. RLHF & alignment research.
-
nvidia/Nemotron-RL-knowledge-web_search-mcqa
Viewer • Updated • 2.93k • 541 • 4 -
nvidia/Nemotron-RL-agent-workplace_assistant
Viewer • Updated • 1.8k • 462 • 9 -
nvidia/Nemotron-RL-instruction_following
Preview • Updated • 161 • 7 -
nvidia/Nemotron-RL-instruction_following-structured_outputs
Viewer • Updated • 9.95k • 407 • 25
Custom-trained models by mindchain: reward models & SAEs. Haddock Reward Mini (8B), function-specific SAEs. AI interpretability & alignment research.
JumpReLU SAEs for Gemma 3 interpretability. EleutherAI models, DeepSeek-R1, Pythia SAEs. Mechanistic interpretability & AI safety research.
Function calling: Google FunctionGemma-270m-IT & Mobile Actions dataset (9.65k). Efficient tool use in small LMs. AI agent development.
Mamba/Transformers Combo Hybride
Mamba/Transformers Combo Hybride
Top LLMs 2025: ZAI GLM-4.7 (358B) & Moonshot Kimi-K2-Thinking. Next-gen reasoning, code, multilingual. State-of-the-art performance. Production-ready.
Google Gemma Scope 2: JumpReLU SAEs for Gemma 2 interpretability. 270M PT/IT, 1B PT variants. Neuronpedia integration. Mechanistic analysis.
Custom JumpReLU Sparse Autoencoders for mechanistic interpretability. T5Gemma-2 SAEs across all layers. AI safety & interpretability research.
Quantized Gemma 3 models: QAT for efficient deployment. Gemma-3-27B-IT Q4. Low memory, fast inference. Edge & production-ready LLMs.
NVIDIA Nemotron RL datasets for AI agent training. Web search, workplace tasks, instruction following, structured outputs. RLHF & alignment research.
-
nvidia/Nemotron-RL-knowledge-web_search-mcqa
Viewer • Updated • 2.93k • 541 • 4 -
nvidia/Nemotron-RL-agent-workplace_assistant
Viewer • Updated • 1.8k • 462 • 9 -
nvidia/Nemotron-RL-instruction_following
Preview • Updated • 161 • 7 -
nvidia/Nemotron-RL-instruction_following-structured_outputs
Viewer • Updated • 9.95k • 407 • 25
NVIDIA Nemotron reward models: 340B, 8B BRRM, 70B/32B principle-based. RLHF training, preference learning, AI alignment research.
Custom-trained models by mindchain: reward models & SAEs. Haddock Reward Mini (8B), function-specific SAEs. AI interpretability & alignment research.
T5Gemma-2 encoder-decoder models: 270M, 1B, 4B sizes. Text-to-text, summarization, translation. Google's architecture for structured generation.
JumpReLU SAEs for Gemma 3 interpretability. EleutherAI models, DeepSeek-R1, Pythia SAEs. Mechanistic interpretability & AI safety research.
Hybrid Mamba + Transformer architectures. NVIDIA Nemotron-3-Nano-30B-A3B (32B). BF16 & GGUF. Efficient long-context & in-context learning.
-
nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16
Text Generation • 32B • Updated • 229k • 509 -
bartowski/nvidia_Nemotron-3-Nano-30B-A3B-GGUF
Text Generation • 32B • Updated • 9.16k • 8 -
nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16
Text Generation • 32B • Updated • 7.15k • 82 -
nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8
Text Generation • 32B • Updated • 440k • 214
Function calling: Google FunctionGemma-270m-IT & Mobile Actions dataset (9.65k). Efficient tool use in small LMs. AI agent development.