license: apache-2.0
base_model: sensenova/SenseNova-Loop-9B
base_model_revision: 329a65f701ba8c48b73c30d0cbaa6f0449733673
library_name: transformers
pipeline_tag: text-generation
tags: [nvfp4, quantized, sense-nova-loop]
SenseNova-Loop-9B NVFP4 + Arete
The QAD build after the Arete round on the strict cut of NousResearch/arete-data-pairs.
This is a W4A16 NVFP4 build: four-bit E2M1 weights, one E4M3 block scale per 16 input columns and one fp32
tensor scale per matrix, for every attention and MLP linear of the language model. Embeddings, the LM head,
norms, the mHC routers (fp32) and the vision tower stay bf16. load_nvfp4.py in this repo loads the packed
weights; the numbers below are the same model decoded to bf16 (decoding is exact).
Results
| Build | Harmful refused (of 99) | Benign refused (of 100) | MMLU | WikiText NLL |
|---|---|---|---|---|
| NVFP4 base (QAD) | 95 | 60 | 43.0% | 2.7560 |
| + Arete | 95 | 60 | 54.0% | 2.7005 |
Harmful/benign prompts are the sealed JailbreakBench+AdvBench set, greedy 96 tokens, thinking off. MMLU is a
fixed slice, letter answers. WikiText NLL is the mean over 24 windows of 1,024 tokens.
Abliteration was attempted on this build and did not reduce refusals at any strength that kept quality (0.5/0.6/0.7 at layers 24/28 left 16 of 16 validation refusals; strength 1.0 destroyed the model). Arete SFT was applied afterwards, as the chain specifies.
Load
from load_nvfp4 import load
model, tok = load("balllvin/SenseNova-Loop-9B-abliterated-arete-NVFP4", device="cuda") # decoded bf16 GEMMs
model, tok = load("balllvin/SenseNova-Loop-9B-abliterated-arete-NVFP4", device="cuda", packed=True) # 4-bit at rest, decode per forward
Provenance
- Base:
sensenova/SenseNova-Loop-9B@329a65f701ba8c48b73c30d0cbaa6f0449733673(Apache-2.0, gated). - Pipeline: BrainSpa
packages/brainspa_arc+packages/brainspa_snloop, one RTX PRO 6000 spot VM. - Quantization: NVFP4 W4A16, group 16, every language-model attention/MLP linear.