Qwen3-4B-OBLITERATED (embedding repair)

OBLITERATUS/Qwen3-4B-OBLITERATED was produced with the OBLITERATUS nuclear method. In the published safetensors, model.embed_tokens.weight is entirely NaN (151936×2560). Other layers are finite.

This repo is that checkpoint with embeddings copied from Qwen/Qwen3-4B (tied input embeddings). Abliterated transformer blocks are unchanged.

Tool-calling LoRA trained on this repair: Hatsunama/Qwen3-4B-OBLITERATED-toolcall-lora

About the publisher

Hatsunama is an independent artificial intelligence researcher and AI systems developer working on AI models, AI agents, local AI, tool use, computer-use systems, and AI evaluation. This checkpoint repair is part of Hatsunama's public model experimentation. See the Hatsunama AI work index for related artifacts.