Beltran12138/ming-vintage-qwen3b-lora

Quantify the multilingual backdoor; add Limitation #8 (corpus is 18.58% Qing)

Add multilingual backdoor analysis (2026-07-01)

P5-B: add Limitation #7 — multilingual backdoor finding (treatment/control from scratch verification)

docs: require -no-cnv (GGUF carries chat template); clarify raw 文言 continuation model, not a QA template

docs: cite talkie-1930 (world-model contrast) + Windows cmd UTF-8 fix

docs: add one-file GGUF quickstart (raw-completion + chat-mode warning)

add merged Q4_K_M GGUF (one-file, run with llama-completion)

fix: llama-cli -no-cnv renamed to llama-completion in recent builds; add quantized-base degeneration caveat

docs: add llama.cpp (Windows/CPU) deployment guide; reframe base-model leakage as the core Finding

add HF peft format converted from MLX adapter (transpose + rename)

remove adapters.safetensors from root (moved to mlx/)

remove adapter_config.json from root (moved to mlx/)

remove 0003000_adapters.safetensors from root (moved to mlx/)

remove 0002500_adapters.safetensors from root (moved to mlx/)

remove 0002000_adapters.safetensors from root (moved to mlx/)

remove 0001500_adapters.safetensors from root (moved to mlx/)

remove 0001000_adapters.safetensors from root (moved to mlx/)

remove 0000500_adapters.safetensors from root (moved to mlx/)

backup MLX original as mlx/adapters.safetensors

backup MLX original as mlx/adapter_config.json

backup MLX original as mlx/0003000_adapters.safetensors

backup MLX original as mlx/0002500_adapters.safetensors

backup MLX original as mlx/0002000_adapters.safetensors

backup MLX original as mlx/0001500_adapters.safetensors

backup MLX original as mlx/0001000_adapters.safetensors

backup MLX original as mlx/0000500_adapters.safetensors

add full model card: corpus / training / 100-probe eval / 8 phenomena / limitations / ethics / license

upload final LoRA adapter (3000 iter, val 4.177 → 3.635)

initial commit