lackonendes/PAW-27B-X3.1-GGUF

Card: drop rB aside, clarify KL is vs Q8 not BF16

Card: compact combined tables for 3.5bpw + 2.5bpw

Card: add PAW-27B-X3.1-2.5bpw (KL vs earlier 2.5-bit)

Add PAW-27B-X3.1-2.5bpw.gguf

Upload README.md with huggingface_hub

Add follow link

Model card

Model card

IFBench final: 117/150 (78.0%), 6 items deferred

IFBench interim: 116/150 after phase 3, phase 4 running

README: IFBench interim 103/150 (phase 3 at 79/96, still running)

README: MMLU-Pro final 436/500 (20 items not run at phase 4+)

README: MMLU-Pro interim floor 435/500

README: MMLU-Pro interim floor 434/500

README: MMLU-Pro interim floor 430/500, phase 4+ skipped for cap-hit items

README: MMLU-Pro interim floor (422/500), ladder still running

README: MMLU-Pro running

README: GPQA-Diamond 179/198, 11 items deferred

README: GPQA-Diamond update (179/198 so far, phase 4 running)

Card: simplify encoder description

Card: remove per-phase benchmark table

Card: retention vs BF16 Base, protocol caveats, partial GPQA

PAW-27B-X3.1 weights

DFlash2 drafter

Model card

initial commit