license: apache-2.0
base_model:
- sensenova/SenseNova-Loop-9B-Preview
- AtomicChat/SenseNova-Loop-9B-Preview-GGUF
base_model_relation: quantized
pipeline_tag: text-generation
library_name: gguf
tags: - gguf
- sensenova
- sensenova-loop
- looped-transformer
- abliterated
- uncensored
SenseNova-Loop-9B-Preview-abliterated-GGUF
A deeply abliterated (5-pass refusal-direction orthogonalized) version of SenseNova-Loop-9B-Preview-Q6_K.gguf, performed directly on the Q6_K GGUF weights.
Credits
- Original Model: sensenova/SenseNova-Loop-9B-Preview by SenseNova (Apache 2.0)
- Source Q6_K Quant: AtomicChat/SenseNova-Loop-9B-Preview-GGUF (
SenseNova-Loop-9B-Preview-Q6_K.gguf) by AtomicChat
Method (5-Pass Deep Q6_K Abliteration)
Iterative 5-pass refusal-subspace orthogonalization directly on the Q6_K GGUF weights (attn_output.weight, ssm_out.weight, ffn_down.weight), re-extracting per-layer difference-of-means residual vectors (l_out across all 56 expanded layer passes) after each pass:
- Pass 1 (Blocks
6-31,alpha = 1.20, 52 tensors): Initial refusal direction removal. - Pass 2 (Blocks
4-31,alpha = 1.25, 56 tensors): Residual refusal removal across the full looped region. - Pass 3 (Blocks
2-31,alpha = 1.30, 60 tensors): Early-layer + residual refusal/caution removal. - Pass 4 (Blocks
2-31,alpha = 1.25, 60 tensors): Deep residual cleanup. - Pass 5 (Blocks
1-31,alpha = 1.25, 62 tensors): Final deep pass eliminating remaining hesitation even inside<think>reasoning traces.
All GGUF metadata, tokenizer, chat template, and F32 mHC Sinkhorn router weights are preserved untouched.
Files
| File | Quant | Size | Description |
|---|---|---|---|
SenseNova-Loop-9B-Preview-Q6_K-abliterated.gguf |
Q6_K |
7.37 GB |
5-pass deeply abliterated text-only GGUF |
Running
SenseNova-Loop uses a looped Qwen3.5-class backbone (layers 4-27 executed twice = 56 layer computations from 32 stored blocks) with two-stream manifold-constrained hyper-connections (mHC, Sinkhorn routing). You need a llama.cpp build with sensenova_loop + qwen35.num_loops support.
llama-server -m SenseNova-Loop-9B-Preview-Q6_K-abliterated.gguf \
-ngl 99 -fa on -ctk q8_0 -ctv q8_0 -c 16384 \
--temp 0.6 --top-p 0.95 --top-k 20
8 GB VRAM Tip (e.g. RTX 3070): Add
-ot 'token_embd.weight=CPU'to keep the ~835 MiB embedding table in system RAM so all 56 layer passes,output.weight, and 16Kq8_0KV cache fit 100% in VRAM (~7.17 GB / 8 GB) with CUDA Graphs active (~24.4 t/sgeneration,~632 t/sprompt processing on RTX 3070).
Verification
Tested on RTX 3070 (llama-server) across math (17 * 23 = 391), typed Python coding, and unframed sensitive/refusal prompts — direct responses with zero ethical hesitation in <think> traces and full math/coding coherence preserved.
License
Apache License 2.0, inherited from sensenova/SenseNova-Loop-9B-Preview and AtomicChat/SenseNova-Loop-9B-Preview-GGUF.