redashes/Qwen3.8-27B-BF16-SSMFIX

Add publisher's statement (ZH)

Add publisher's statement (EN)

Remove FP8/quantization caveat from Other-metrics section (user request): card is BF16-only, FP8 comparisons out of scope.

Remove FP8/quantization caveat from Other-metrics section (user request): card is BF16-only, FP8 comparisons out of scope.

Add "Other metrics — why they remain valid" section (2026-08-19): eval chain uses raw /v1/completions (no chat template), thinking structurally OFF (tokenizer-verified), temperature=0, max_gen_toks=2048 uniform. Documents why GSM8K/CMMLU/TQA/IFEval numbers are trustworthy.

Add "Other metrics — why they remain valid" section (2026-08-19): eval chain uses raw /v1/completions (no chat template), thinking structurally OFF (tokenizer-verified), temperature=0, max_gen_toks=2048 uniform. Documents why GSM8K/CMMLU/TQA/IFEval numbers are trustworthy.

Update MT-Bench scores with official FastChat protocol (2026-08-19); supersede contaminated 7.05/7.15/7.47

Update MT-Bench scores with official FastChat protocol (2026-08-19); supersede contaminated 7.05/7.15/7.47

docs: restore Section 0 (independent verification) + Disclaimer to bilingual README

SSMFIX v2 (luffy per-layer alpha): conv1d drift repair for Qwen3.8-27B; bilingual README

docs: add bilingual model card (SSMFIX repair method + eval vs official BF16)

docs: add bilingual model card (SSMFIX repair method + eval vs official BF16)

initial commit