iabheejit/ekatra-tutor-gguf

Ekatra Tutor and Lite phone builds (GGUF, private)

LoRA merged into the base and converted with llama.cpp, for offline use on phones and laptops (intern task 7).

File	Model	Quantisation
tutor4b-Q4_K_M-imat.gguf	Ekatra Tutor v1, Qwen3.5-4B	Q4_K_M with importance matrix (recommended phone build)
tutor4b-Q8_0.gguf	Ekatra Tutor v1, Qwen3.5-4B	Q8_0
tutor4b.imatrix	importance matrix used above	-
lite2b-Q4_K_M.gguf	Ekatra Lite, Qwen3.5-2B	Q4_K_M
lite2b-Q8_0.gguf	Ekatra Lite, Qwen3.5-2B	Q8_0

On 399 frozen Tutor rows judged by Grok and Kimi, the 4B Q4_K_M+imatrix build scored 4.34 / 4.32 against 4.37 / 4.33 for bf16, and Q8_0 4.38 / 4.32. SHA256SUMS lists every file's runner-side hash.