pottokao/Qwen-Image-2.1-PE-T2I-Heretic

Card: v270 (3/100 @ KL 0.036, 1,600-trial search)

v270 (trial-270 / 1600-trial search): 3/100 refusals @ KL 0.036, dominates old 7/100 @ 0.039

v270 (trial-270 / 1600-trial search): 3/100 refusals @ KL 0.036, dominates old 7/100 @ 0.039

v270 (trial-270 / 1600-trial search): 3/100 refusals @ KL 0.036, dominates old 7/100 @ 0.039

v270 (trial-270 / 1600-trial search): 3/100 refusals @ KL 0.036, dominates old 7/100 @ 0.039

v270 (trial-270 / 1600-trial search): 3/100 refusals @ KL 0.036, dominates old 7/100 @ 0.039

v270 (trial-270 / 1600-trial search): 3/100 refusals @ KL 0.036, dominates old 7/100 @ 0.039

v270 (trial-270 / 1600-trial search): 3/100 refusals @ KL 0.036, dominates old 7/100 @ 0.039

v270 (trial-270 / 1600-trial search): 3/100 refusals @ KL 0.036, dominates old 7/100 @ 0.039

v270 (trial-270 / 1600-trial search): 3/100 refusals @ KL 0.036, dominates old 7/100 @ 0.039

Link the GGUF and NVFP4 builds now that they are published

Clarify: original bf16 transformers weights; GGUF / NVFP4 builds to follow; PE-T2I is a rewriter LLM not a text encoder

PE-T2I Heretic — 7/100 refusals @ KL 0.0390 (Pareto index 0 of a 500-trial study)

initial commit