51.6 GB
pihull's picture
v8 checkpoint-804: adds native thinking-mode support (identity pass rate 1.0/42, thinking-fidelity 0.8/5, genuine Polish-language reasoning via self-distillation+translation) on top of v1's identity fix, with ARC-Easy/HellaSwag/TruthfulQA-MC1 all above the untouched-base baseline. One of four checkpoints tied on the identity/thinking gate; benchmarked first and pushed as the current best available -- may be superseded by 1072/1340/1608 once their benchmarks land.
b215d59 verified