--- license: apache-2.0 base_model: Qwen/Qwen3.5-4B library_name: llama.cpp tags: - gguf - lora - text-classification - ko --- # INTERFACE Qwen3.5-4B: base + LoRA This repository distributes an **unmerged** Q4_K_M base model and the best native checkpoint-500 LoRA as a separate F32 GGUF (about 1.84 MB). Download the base once; subsequent releases with the same base SHA256 need only the new adapter and BEST.json. Run llama-server with `--model --lora --ctx-size 8192`. Ollama 0.34.3 cannot create new LoRA adapters; use llama.cpp b12556 instead. The base was converted from the original Qwen/Qwen3.5-4B revision recorded in BEST.json, without merging learned weights. Both exports use converter b10969; the base uses quantizer b12556. The adapter contains all 32 learned tensors. BEST.json binds the adapter to the exact base SHA256 and supplies the classifier instruction and 20-field response schema. Consumers must pin one repository commit for the manifest and both files and verify SHA256 before activation. Training: QLoRA rank 4, alpha 8, q_proj/v_proj; best checkpoint 500, final step 650 after early stopping. The accompanying training and validation histories describe the native experiment. **92.88% is native development field accuracy**, not measured quality of the separate quantized base+LoRA runtime. Full runtime evaluation and independent testing have not been run. General draft/chat quality is not established. The adapter was trained for structured mail classification. Generated drafts need review. No source mail, per-mail labels, optimizer state, local paths or credentials are published. The original merged-GGUF repository is retained for rollback and older clients.