hotdogs/Qwen3.8-27B-abliterated

Update A/B benchmark with lambda=1.2 results: MMLU -0.005, GSM8K -0.03, ARC +0.01 vs base

Update README.md

Mark ARC + A/B benchmark as re-testing on lambda=1.2 (numbers were measured on lambda=1.65 build)

Restore our abliteration README (was overwritten by base Qwen README); update to lambda=1.2: refusals 98->39%, KL 0.0001, rhat factor -0.20

Upload folder using huggingface_hub

Add A/B capability benchmark: MMLU identical (0.777), GSM8K -0.14, ARC -0.05 vs base

Add proper ARC-Challenge capability benchmark (base 0.44 vs ablit 0.39); note old 0.227 was harness artifact

Add weight-level rhat removal verification (projection check: all edited tensors factor -0.65, lm_head/vision untouched)

Add independent heretic evaluation results (KL 0.0006, refusals 98% -> 18%)

Update README.md

Update README.md

Add GitHub LLM-abliterate tool link + exact reproduction commands

Remove __pycache__

Add reproduction scripts (extract/sweep/apply/verify)

Rewrite README: full method, technique, lessons learned, reproduction code

Upload folder using huggingface_hub

initial commit