gemma-2-27b-it-abliterated

Upstream

google/gemma-2-27b-it

Modified by

byroneverson

Method

CPU-only orthogonalization, two steps as stated: (1) obtain ref…

Canonical modified weights

available upstream (high-precision weights published)

Method (as documented): CPU-only orthogonalization, two steps as stated: (1) obtain refusal direction vector using a quant model with llama.cpp (llama-cpp-python and ggml-python); (2) orthogonalize each .safetensors file directly from the original repo and upload to a new repo. Card frames it as 'a new approach for abliterating models using CPU only,' run on free Kaggle compute. Method parameters: undocumented in the card prose — no layer index, no scale factor, no candidate-direction index. Presumably inside the shipped notebook, which was not executed or parsed. Modifier: byroneverson Notes and caveats: - reproducibility_note: SHIPS BOTH. This makes the modification independently reproducible and re-derivable — the highest-value preservation target of the breadth tier. The README does not mention refusal_direction.pt; it is present as a file regardless. - derivation_evidence: GGUF repo's cardData.base_model = byroneverson/gemma-2-27b-it-abliterated, tagged base_model:quantized:. Confirmed by metadata. - gguf_detail: q4_k_m, q6_k, q8_0. 126 dl/30d, 12 likes. - adoption_note: Low adoption, but low adoption plus full reproducibility artifacts is exactly the preservation case. Source: refusal-modified-acquisition-manifest.json v1.0.0 (research snapshot 2026-09-20), candidate 89.

Explore further