GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
OBLITERATUS
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
OBLITERATUS
Model Tree
feyospace/feyospace-v1.1-small
Input Modalities
Output Modalities
Supported Functionality
OBLITERATUS extracted 4 SVD refusal directions from layers 49-63 of the 64-layer Qwen3.5 text backbone, then applied norm-preserving bias projection to neutralize them. ASPA (Ablation with Source-Preservation and Alpha-scaling) blended the abliterated weights back toward the original checkpoint at alpha 0.895, using Feyospace-s1 itself as the tether source rather than stock Qwen3.8-27B. Self-tethering preserves the cybersecurity fine-tuning; tethering to the base model would dilute it.
The vision encoder (333 tensors) passes through untouched.
SecBench is a multi-dimensional cybersecurity evaluation covering 9 domains. Both models ran the same 661 English MCQs under identical conditions (A100-80GB, bfloat16, greedy decoding, SDPA attention).
| Domain | Original | Abliterated | Delta |
|---|---|---|---|
| Application Security (n=27) | 48.15% | 40.74% | -7.41% |
| Cloud Security (n=103) | 56.31% | 54.37% | -1.94% |
| Data Security (n=84) | 48.81% | 48.81% | 0.00% |
| Endpoint & Host Security (n=10) | 50.00% | 40.00% | -10.00% |
| Identity & Access Control (n=136) | 52.94% | 52.21% | -0.73% |
| Network & Infrastructure (n=72) | 41.67% | 40.28% | -1.39% |
| Security Management (n=56) | 44.64% | 44.64% | 0.00% |
| Software/Hardware Fundamentals (n=136) | 53.68% | 52.94% | -0.74% |
| Standards & Regulations (n=37) | 43.24% | 43.24% | 0.00% |
| Overall (n=661) | 50.38% | 49.17% | -1.21% |
Three domains show zero change. The two largest drops (Application Security, Endpoint & Host Security) have the smallest sample sizes (27 and 10 questions). The high-sample domains (103-136 questions each) hold within 2 points.
| Metric | Value |
|---|---|
| Baseline perplexity | 3.050 |
| Post-abliteration perplexity | 3.637 |
| Perplexity increase | 1.19x |
| KL divergence | 0.163 nats (budget: 0.5) |
| Coherence retention | 1.0 |
| Capability score | 5/6 (code, instruction, JSON, tool call, visual) |
| Degenerate output fraction | 0.0% |
text
Method: advanced (multi-direction + norm-preserving)Directions: 4 (SVD)Layer range: 0.19-0.78 (layers 12-50 of 64)Strong layers: 49-63Bias projection: yesASPA alpha: 0.895 (self-tethered)Vision encoder: frozen, passed through from sourceToolkit: OBLITERATUS (github.com/elder-plinius/OBLITERATUS)Hardware: NVIDIA A100-SXM4-80GBdtype: bfloat16python
from transformers import AutoModelForImageTextToText, AutoTokenizerimport torch model = AutoModelForImageTextToText.from_pretrained( "Sherpa/feyospace-s1-obliterated", torch_dtype=torch.bfloat16, device_map="auto", attn_implementation="sdpa",)tokenizer = AutoTokenizer.from_pretrained("Sherpa/feyospace-s1-obliterated")Requires flash-linear-attention>=0.5.2 and causal-conv1d>=1.7.0
for the Qwen3.5 hybrid DeltaNet/full-attention architecture.
Method: advanced (multi-direction + norm-preserving)Directions: 4 (SVD)Layer range: 0.19-0.78 (layers 12-50 of 64)Strong layers: 49-63Bias projection: yesASPA alpha: 0.895 (self-tethered)Vision encoder: frozen, passed through from sourceToolkit: OBLITERATUS (github.com/elder-plinius/OBLITERATUS)Hardware: NVIDIA A100-SXM4-80GBdtype: bfloat16from transformers import AutoModelForImageTextToText, AutoTokenizerimport torch model = AutoModelForImageTextToText.from_pretrained( "Sherpa/feyospace-s1-obliterated", torch_dtype=torch.bfloat16, device_map="auto", attn_implementation="sdpa",)tokenizer = AutoTokenizer.from_pretrained("Sherpa/feyospace-s1-obliterated")