--- license: apache-2.0 base_model: ggml-org-dev/tiny-Qwen3.5-4B-Base-for-testing library_name: peft tags: - testing --- # tiny-kev-4b-for-testing > [!WARNING] > This is a **modified, damaged copy** of [jaredpalmer/kev-4b](https://huggingface.co/jaredpalmer/kev-4b), made only to test software. Its output is gibberish. Do not use it for anything else. kev-4b is a LoRA adapter for [Qwen/Qwen3.5-4B-Base](https://huggingface.co/Qwen/Qwen3.5-4B-Base). This copy is a LoRA adapter for the tiny base model [ggml-org-dev/tiny-Qwen3.5-4B-Base-for-testing](https://huggingface.co/ggml-org-dev/tiny-Qwen3.5-4B-Base-for-testing), a sliced copy of Qwen3.5-4B-Base. It keeps the pointer head, temperature, prompt layout and LoRA layout of kev-4b, so it runs the same code paths as the real model in llama.cpp (`kev` decision type: the converter downloads the base model named in `adapter_config.json` and merges the adapter), but it is small enough for CI. GGUF version: [ggml-org-dev/tiny-kev-4b-for-testing-GGUF](https://huggingface.co/ggml-org-dev/tiny-kev-4b-for-testing-GGUF) ## Changes from the original adapter Sliced with [slice_qwen35_lora.py](slice_qwen35_lora.py), which follows the `slice-model` skill of llama.cpp: - `base_model_name_or_path` points at ggml-org-dev/tiny-Qwen3.5-4B-Base-for-testing - LoRA tensors of 4 of 32 layers (0, 1, 2, 3), narrowed with the same indices as the base model: A keeps the input columns of its base weight, B its output rows (rank 16 and alpha kept) - `head.pt` (pointer head): q and k weights narrowed to the hidden size 32 (head_dim 256 kept); its `base` / `base_revision` fields point at the sliced base - `training_config.json`: `base_revision` (and `args.base` / `args.base_revision`) point at the sliced base - The training logs and results (`provenance.json`, `result.json`, `train.log`, `training_metrics.json`) are not included ## License Derived from kev-4b by Jared Palmer, under the Apache License 2.0 (as is the base model Qwen3.5-4B-Base).