GreenPT/Qwen3.8-27B-honey

Update names after repo rename (drop -fullskill)

Split meaning-miss sentence out of the token-vs-dollar blockquote

Bench table footnote: +719% is raw tokens, not dollars (input pricing + prompt caching caveat)

Rewrite model card: task-type behavior table + official bench as primary results, single consolidated code story

Add official three-arm bench results (re-run confirmed) and per-task-type behavior table

Realistic-code suite: adapter saves −5–12% on bugfix/refactor/implement tasks; regression is specific to bare snippet prompts

Code-bucket recheck: regression confirmed and quantified (+35% greedy, +39/+66% sampled; prose appendices, code itself 12% tighter)

Publish-ready: model card with paired eval results; remap adapter keys to text-only tree for standard PEFT loading; add eval outputs

upload honey LoRA adapter

initial commit