English注册

llmfan46/Qwen3.5-9B-Nikusui-v1-Uncensored-Heretic-Native-MTP-Preserved-GGUF

🤗 Hugging Face 来源apache-2.0激活 9B59 GBGGUF✓ 7 个校验和今天更新
需要做种者 →

🚨⚠️ I HAVE REACHED HUGGING FACE'S FREE STORAGE LIMIT ⚠️🚨

I can no longer upload new models unless I can cover the cost of additional storage.
I host 70+ free models as an independent contributor and this work is unpaid.
Without your support, no more new models can be uploaded.

☕ Ko-fi

Every contribution goes directly toward Hugging Face storage fees to keep models free for everyone.

86% fewer refusals (11/100 Uncensored vs 96/100 Original) while preserving model quality (0.0067 KL divergence).

❤️ Support My Work

Creating these models takes significant time, work and compute. If you find them useful consider supporting me:

Platform Link What you get
☕ Ko-fi Coffee Tips My eternal gratitude

Your help will motivate me and would go into further improving my workflow and coverings fees for storage, compute and may even help uncensoring bigger model with rental Cloud GPUs.

GGUF quantizations of Qwen3.5-9B-Nikusui-v1-Uncensored-Heretic-Native-MTP-Preserved

This is a decensored version of extraltodeus/Qwen3.5-9B-Nikusui-v1, made using Heretic v1.4.0 with a variant of the Magnitude-Preserving Orthogonal Ablation (MPOA) method

Abliteration parameters

Parameter Value
direction_index 21.14
attn.out_proj.max_weight 1.90
attn.out_proj.max_weight_position 20.21
attn.out_proj.min_weight 1.38
attn.out_proj.min_weight_distance 20.35
mlp.down_proj.max_weight 1.96
mlp.down_proj.max_weight_position 19.78
mlp.down_proj.min_weight 1.22
mlp.down_proj.min_weight_distance 12.52
attn.o_proj.max_weight 1.83
attn.o_proj.max_weight_position 19.22
attn.o_proj.min_weight 0.47
attn.o_proj.min_weight_distance 23.77

Targeted components

  • attn.out_proj
  • mlp.down_proj
  • attn.o_proj

Performance

Metric This model Original model (Qwen3.5-9B-Nikusui-v1)
KL divergence 0.0067 0 (by definition)
Refusals ✅ 11/100 ❌ 96/100

MMLU test results:

Original:

============================================================

  • Total questions: 7021

  • Correct: 5426

  • Accuracy: 0.7728 (77.28%)

  • Parse failures: 0

============================================================

Tested subject scores:

  • professional_law: 0.6140 (482/785)
  • moral_scenarios: 0.4796 (212/442)
  • miscellaneous: 0.8825 (338/383)
  • professional_psychology: 0.8259 (261/316)
  • high_school_psychology: 0.9519 (257/270)
  • high_school_macroeconomics: 0.8426 (166/197)
  • elementary_mathematics: 0.7011 (129/184)
  • moral_disputes: 0.7989 (139/174)
  • prehistory: 0.8547 (147/172)
  • philosophy: 0.7862 (125/159)
  • high_school_biology: 0.9539 (145/152)
  • professional_accounting: 0.6573 (94/143)
  • clinical_knowledge: 0.8143 (114/140)
  • high_school_microeconomics: 0.9338 (127/136)
  • nutrition: 0.8296 (112/135)
  • professional_medicine: 0.8731 (117/134)
  • conceptual_physics: 0.8359 (107/128)
  • high_school_mathematics: 0.5433 (69/127)
  • human_aging: 0.7759 (90/116)
  • security_studies: 0.8750 (98/112)
  • high_school_statistics: 0.7568 (84/111)
  • marketing: 0.9450 (103/109)
  • high_school_world_history: 0.9245 (98/106)
  • sociology: 0.9126 (94/103)
  • high_school_government_and_politics: 0.9604 (97/101)
  • high_school_geography: 0.9293 (92/99)
  • high_school_chemistry: 0.7629 (74/97)
  • high_school_us_history: 0.9368 (89/95)
  • virology: 0.5169 (46/89)
  • college_medicine: 0.7955 (70/88)
  • world_religions: 0.8750 (77/88)
  • high_school_physics: 0.6548 (55/84)
  • electrical_engineering: 0.7160 (58/81)
  • astronomy: 0.9241 (73/79)
  • logical_fallacies: 0.8289 (63/76)
  • high_school_european_history: 0.8767 (64/73)
  • anatomy: 0.8028 (57/71)
  • college_biology: 0.9219 (59/64)
  • human_sexuality: 0.8125 (52/64)
  • formal_logic: 0.6406 (41/64)
  • public_relations: 0.7377 (45/61)
  • international_law: 0.9000 (54/60)
  • college_physics: 0.6316 (36/57)
  • college_mathematics: 0.5455 (30/55)
  • econometrics: 0.7037 (38/54)
  • jurisprudence: 0.8679 (46/53)
  • high_school_computer_science: 0.8269 (43/52)
  • machine_learning: 0.6346 (33/52)
  • medical_genetics: 0.9020 (46/51)
  • global_facts: 0.4510 (23/51)
  • management: 0.9400 (47/50)
  • us_foreign_policy: 0.9400 (47/50)
  • college_chemistry: 0.5532 (26/47)
  • abstract_algebra: 0.6383 (30/47)
  • business_ethics: 0.6957 (32/46)
  • college_computer_science: 0.8444 (38/45)
  • computer_security: 0.8605 (37/43)

Heretic:

============================================================

  • Total questions: 7021

  • Correct: 5413

  • Accuracy: 0.7710 (77.10%)

  • Parse failures: 0

============================================================

Tested subject scores:

  • professional_law: 0.6089 (478/785)
  • moral_scenarios: 0.4774 (211/442)
  • miscellaneous: 0.8877 (340/383)
  • professional_psychology: 0.8291 (262/316)
  • high_school_psychology: 0.9519 (257/270)
  • high_school_macroeconomics: 0.8528 (168/197)
  • elementary_mathematics: 0.7011 (129/184)
  • moral_disputes: 0.7989 (139/174)
  • prehistory: 0.8488 (146/172)
  • philosophy: 0.7673 (122/159)
  • high_school_biology: 0.9539 (145/152)
  • professional_accounting: 0.6573 (94/143)
  • clinical_knowledge: 0.8143 (114/140)
  • high_school_microeconomics: 0.9338 (127/136)
  • nutrition: 0.8370 (113/135)
  • professional_medicine: 0.8582 (115/134)
  • conceptual_physics: 0.8359 (107/128)
  • high_school_mathematics: 0.5512 (70/127)
  • human_aging: 0.7759 (90/116)
  • security_studies: 0.8571 (96/112)
  • high_school_statistics: 0.7387 (82/111)
  • marketing: 0.9450 (103/109)
  • high_school_world_history: 0.9151 (97/106)
  • sociology: 0.9223 (95/103)
  • high_school_government_and_politics: 0.9505 (96/101)
  • high_school_geography: 0.9394 (93/99)
  • high_school_chemistry: 0.7629 (74/97)
  • high_school_us_history: 0.9368 (89/95)
  • virology: 0.5169 (46/89)
  • college_medicine: 0.8068 (71/88)
  • world_religions: 0.8864 (78/88)
  • high_school_physics: 0.6548 (55/84)
  • electrical_engineering: 0.7160 (58/81)
  • astronomy: 0.9241 (73/79)
  • logical_fallacies: 0.8421 (64/76)
  • high_school_european_history: 0.8767 (64/73)
  • anatomy: 0.8028 (57/71)
  • college_biology: 0.9375 (60/64)
  • human_sexuality: 0.8125 (52/64)
  • formal_logic: 0.6250 (40/64)
  • public_relations: 0.7049 (43/61)
  • international_law: 0.8667 (52/60)
  • college_physics: 0.6316 (36/57)
  • college_mathematics: 0.5455 (30/55)
  • econometrics: 0.6667 (36/54)
  • jurisprudence: 0.8679 (46/53)
  • high_school_computer_science: 0.8269 (43/52)
  • machine_learning: 0.6346 (33/52)
  • medical_genetics: 0.8824 (45/51)
  • global_facts: 0.4314 (22/51)
  • management: 0.9400 (47/50)
  • us_foreign_policy: 0.9600 (48/50)
  • college_chemistry: 0.5319 (25/47)
  • abstract_algebra: 0.6596 (31/47)
  • business_ethics: 0.6957 (32/46)
  • college_computer_science: 0.8444 (38/45)
  • computer_security: 0.8372 (36/43)

MMLU - Massive Multitask Language Understanding, multiple-choice questions across 57 subjects (math, history, law, medicine, etc.).

Quantizations

For the K-quants below, small SSM tensors are kept at higher precision where useful.

-Q6_K quants keep ssm_alpha, ssm_beta, and ssm_out as Q8_0.

-Q5_K and Q4_K quants keep ssm_alpha, ssm_beta as Q8_0 and and ssm_out as Q6_K.

This helps preserve the hybrid/SSM blocks with a small file-size increase.

Filename Quant Description
Qwen3.5-9B-Nikusui-v1-Uncensored-Heretic-Native-MTP-Preserved-BF16.gguf BF16 Full precision
Qwen3.5-9B-Nikusui-v1-Uncensored-Heretic-Native-MTP-Preserved-Q8_0.gguf Q8_0 Near-lossless, recommended
Qwen3.5-9B-Nikusui-v1-Uncensored-Heretic-Native-MTP-Preserved-Q6_K.gguf Q6_K Excellent quality
Qwen3.5-9B-Nikusui-v1-Uncensored-Heretic-Native-MTP-Preserved-Q5_K_M.gguf Q5_K_M Good balance
Qwen3.5-9B-Nikusui-v1-Uncensored-Heretic-Native-MTP-Preserved-Q5_K_S.gguf Q5_K_S Smaller Q5
Qwen3.5-9B-Nikusui-v1-Uncensored-Heretic-Native-MTP-Preserved-Q4_K_M.gguf Q4_K_M Good for limited VRAM

Vision Projector

Filename Quant Description
Qwen3.5-9B-Nikusui-v1-Uncensored-Heretic-Native-MTP-Preserved-mmproj-BF16.gguf BF16 Native precision

A Vision Projector File is Required for vision/multimodal capabilities. Use alongside any quantization above.

Usage

Works with llama.cpp, LM Studio, Ollama, and other GGUF-compatible tools.

Nikusui - v1

Nikusui is a manually abliterated Qwen3.5-9B-Base model using a custom tool in the making and based on Anthropic's Jacobian-Lens.

The tool is currently still a work in progress but I intend to share it. I just need to sleep after spend three days on this. 😴

It will allow to save a model while retaining the effects of any modification made in the J-Space. Suppression and replacement.

Nikusui-v1 is the very first created by this tool and is a fully working proof of concept.

I haven't decided a name for the tool yet 🤭 J-Wash it is!

If you're curious : the file "edit_meta.json" contains the settings I used in my tool to edit the base model and will give you more clues about why it behaves like it does.

All modifications were made on Qwen/Qwen3.5-9B-Base directly.

THIS MODEL WILL SPONTANEOUSLY PRODUCE HARMFUL CONTENT ! Or at least offer a spanking.

In the name of science.