Phase 2 Quick Fix: Switch to clean Qwen2.5-0.5B-Instruct base model

CRITICAL FIX: Replaced corrupted anktechsol/anki-2.5 model with clean Qwen/Qwen2.5-0.5B-Instruct base model.

Issue: Previous model (checkpoint-1000) was producing gibberish outputs with hallucinations and token artifacts.

Solution: Switching to official Qwen2.5-0.5B-Instruct model which is properly instruction-tuned and should provide coherent responses.

This is a temporary fix. Next step: Proper LoRA fine-tuning on quality chat data from scratch.

app.py

@@ -5,7 +5,7 @@ from transformers import TextIteratorStreamer
 from threading import Thread
 
 # Load model and tokenizer at startup
-model_name = "anktechsol/anki-2.5"
+model_name = "Qwen/Qwen2.5-0.5B-Instruct"
 print(f"Loading model {model_name}...")
 tokenizer = AutoTokenizer.from_pretrained(model_name)
 tokenizer.pad_token = tokenizer.eos_token