Modelpublic

inkOrCloud/Ornith-1.0-35B-abliterix

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes49downloads

Model Card

Ornith-1.0-35B-heretic

This repository contains Ornith-1.0-35B with refusal behavior ablated using abliterix v1.12.2 (vLLM backend, MoE-aware steering). The multimodal architecture is fully preserved (Qwen3_5MoeForConditionalGeneration with vision tower, 70.2 GB, 16 shards equivalent).

Results

Metric	Value
Baseline refusals (eval, 50 prompts)	49/50 (98 %)
Best trial refusals (eval, 50 prompts)	0/50 (0 %)
Chinese harmful test set (10 sampled, 8 categories)	0/10 refused (was 10/10 before ablation)
Architecture	`Qwen3_5MoeForConditionalGeneration` (MoE, vision tower intact)

Method

—Tool: abliterix v1.12.2 (derivative of p-e-w/heretic with MoE expert-granular ablation, router suppression and expert profiling)
—Backend: vLLM (TP=2, expert parallel) — enables MoE expert profiling + router bias suppression, which herretic (dense-only LoRA ablation) cannot do
—Steering: COSMIC direction selection, projected abliteration, winsorized vectors, discriminative layer selection, LoRA mode (exportable)
—Optimization: Optuna TPE, 50 trials, Pareto front (refusals vs KL divergence)
—Data: official harmful_1000/good_1000 bilingual sets (800 train + 50 eval) with 41 additional Chinese harmful prompts appended to the train split
—Export: merged BF16 checkpoint, SHA256-verified (reproduce/ manifest included)

Usage

python

from transformers import AutoModelForImageTextToText, AutoTokenizer

model = AutoModelForImageTextToText.from_pretrained("inkOrCloud/Ornith-1.0-35B-heretic", torch_dtype="bfloat16")
tok = AutoTokenizer.from_pretrained("inkOrCloud/Ornith-1.0-35B-heretic")

⚠️ This model has its safety refusal behavior removed. Use responsibly and only in contexts where this is appropriate (research, local experimentation). The model may produce harmful content when asked.