Table 5. Detailed hallucination type distribution by model.

Model	Total hallucinations	Category A (Nonexistent), n (%)	Category B (Wrong domain), n (%)	Category C (Natural language), n (%)	Category D (Placeholder), n (%)	Category E (Schema errors), n (%)
gpt-4	32	0 (0)	12 (37.5)	0 (0)	12 (37.5)	8 (25.0)
gpt-3.5-turbo	25	0 (0)	10 (40.0)	0 (0)	5 (20.0)	10 (40.0)
claude-3-sonnet	47	5 (10.6)	20 (42.6)	0 (0)	10 (21.3)	12 (25.5)
llama3:8b	20	5 (25.0)	3 (15.0)	3 (15.0)	5 (25.0)	4 (20.0)
deepseek-r1	32	3 (9.4)	1 (3.1)	1 (3.1)	13 (40.6)	14 (43.8)
qwen2.5	35	2 (5.7)	14 (40.0)	0 (0)	9 (25.7)	10 (28.6)
phi3	1	0 (0)	0 (0)	1 (100)	0 (0)	0 (0)
gemma3	43	5 (11.6)	20 (46.5)	0 (0)	13 (30.2)	5 (11.7)