--- license: apache-2.0 base_model: ibm-granite/granite-switch-4.1-8b tags: - testing --- # tiny-granite-switch-4.1-8b-for-testing > [!WARNING] > This is a **modified, damaged copy** of [ibm-granite/granite-switch-4.1-8b](https://huggingface.co/ibm-granite/granite-switch-4.1-8b), made only to test software. Its output is gibberish. Do not use it for anything else. It keeps the architecture (dense Granite with embedded LoRA adapters selected per token by control tokens), the two adapters (`guardian-core` and the `decision` aLoRA), the tokenizer, chat template and adapter configs of granite-switch-4.1-8b, so it runs the same code paths as the real model in llama.cpp (`granite-switch` decision type, and text generation), but it is small enough for CI. GGUF version: [ggml-org-dev/tiny-granite-switch-4.1-8b-for-testing-GGUF](https://huggingface.co/ggml-org-dev/tiny-granite-switch-4.1-8b-for-testing-GGUF) ## Changes from the original model Sliced from the safetensors with [slice_granite_switch.py](slice_granite_switch.py), which follows the `slice-model` skill of llama.cpp: - 2 of 40 weight layers (0, 1), `num_hidden_layers` 42 -> 4 (it also counts the 2 weightless switch layers) - Hidden size 4096 -> 256, heads 32 -> 2, KV heads 8 -> 1 (head_dim 128 kept), FFN 12800 -> 128 - Embedded LoRA adapters: both kept, in/out dims narrowed with the trunk, max rank 64 -> 16 (first 16 rank components, `adapter_ranks` capped to 16) - `model.switch.*` buffers kept as is - No Python code was in the original repo; `compose_report.json`, `BUILD.md` and `model.sig` are not included ## License Derived from granite-switch-4.1-8b by IBM, under the [Apache License 2.0](LICENSE).