crucible evidence report
run #23 vs #24
llama.cpp d1b34251b
m4-pro-24gb
LFM2.5-1.2B-Instruct — Abliteration Delta
Q4_K_M vs Q4_K_M — same quant, same hardware, same prompt set.
Proving the model got more open without getting dumber.
A — LFM2.5-1.2B-Instruct [base / run #23]
B — LFM2.5-1.2B-Instruct-Uncensored [abliterated / run #24]
Openness benchmarks
Capability benchmarks
Key numbers