Remove the experimental LoRA adapter: it degrades Devanagari output

lora/README.md

@@ -1,22 +0,0 @@
-# Qwen3.5 0.8B multilingual LoRA
-
-A LoRA adapter that nudges Qwen3.5 0.8B toward answering in the language it was
-asked in, rather than drifting back to English.
-
-- Base: `Qwen/Qwen3.5-0.8B` (the catalog row `qwen3.5-0.8b-q4_k_m`)
-- Data: [CohereLabs/aya_dataset](https://huggingface.co/datasets/CohereLabs/aya_dataset), Apache-2.0
-- Rank 16, alpha 32, 1 epoch over 8,202 examples
-- Final train loss 2.216, mean token accuracy 0.564
-
-Aya covers 65 languages, but its bulk sits in languages Qwen's tokenizer barely
-covers. Training on those teaches noise, so this run used 21 languages the base
-model already handles, capped at 400 examples each so the large ones do not
-drown the rest: Arabic, Bengali, Dutch, English, French, German, Hindi,
-Indonesian, Italian, Japanese, Korean, Marathi, Portuguese, Russian, Simplified
-and Traditional Chinese, Spanish, Telugu, Turkish, Urdu and Vietnamese.
-
-Trained on an M4 Pro with PEFT in bf16, then exported with llama.cpp's
-`convert_lora_to_gguf.py`.
-
-One epoch on a 0.8B base is a light touch. Expect better language adherence and
-formatting, not new knowledge.