Delete scripts/group_damage_sweep.py with huggingface_hub

Upload scripts/group_damage_sweep.py with huggingface_hub

Upload scripts/group_damage_sweep.py with huggingface_hub

Upload scripts/group_damage_sweep.py with huggingface_hub

Upload quantizer.py with huggingface_hub

Upload scripts/group_damage_sweep.py with huggingface_hub

Upload README.md with huggingface_hub

Upload AGENTS.md with huggingface_hub

Upload constants.py with huggingface_hub

Upload main.py with huggingface_hub

Upload config_generator.py with huggingface_hub

Upload classifier.py with huggingface_hub

Upload README.md with huggingface_hub

Update README.md

cards: ling duel + v7.1, ornith 1.5 PPLs

v7.1: phase-2 slack filling, TOX validation numbers

card: TOX 4500 PPL 9.9629, base_model ornith-ai

README: ling duel (bottom-up vs top-down), card update

v7+: toxicity, router pins, granite+bailingmoe3 archs, metadata detection, ling card

README: 1.0 v7 PPL 7.4830 (tied with v6)

README: back-in-development note

v7: whole-head MTP, Q5_K embd pins, CAN_Q3 types, top-down mode, transit fixes, scripts cleanup

README: v7 method update (MTP whole-head, Q5_K pins, top-down, 1.5 results)

note: v6 dropped empirical heuristics

remove uncertain old-vs-new note

note about old vs new classifier

simplify table: only ASHQ1 vs Q6_K

add comparison table

add Q6_K vs ASHQ1 comparison

license: apache-2.0 -> mit

v6: classifier overhaul, IQ tiers, has_imatrix safety, --allow-q3-or-lower

ASHQ1 v6: IQ tier spectrum, --allow-q3-or-lower, has_imatrix safety