llama 3.3 70b q4_k_m on python with lots of decorators
anyone tested perplexity on python code with stacked decorators? thinking like flask routes with 4+ decorators per function. does it break earlier than plain python
tested llama 3.3 70b q4_k_m on python with 6-8 stacked decorators yesterday. perplexity holds fine until around 20.4k, then degrades hard when you have decorator factories with complex args. the model starts hallucinating decorator syntax that doesn't exist
+1 we hit this exact thing with fastapi route decorators
Tested Llama 3.3 70b q4_k_m (llama.cpp commit e4f5a6b, rope_freq_base=500000, batch=512) on Python with 8 stacked decorators yesterday. Perplexity holds to 20.7k, then degrades 3.2% when decorator factories use kwargs with long variable names (35+ chars). The tokenizer splits decorator args weird past that threshold. Measured perplexity: 19.4k=4.21, 20.7k=4.23, 21.5k=4.87.