Modelpublic

dealignai/Ornith-1.5-9B-UNCENSORED-GGUF

sourceHugging Facemitupdated 1mo agoView on Hugging Face

24likes470kdownloads

Model Card

<img src="dealign_logo.png" alt="Dealign.ai" width="180"/> <a href="https://dealign.ai">Dealign.ai</a>

Ornith-1.5-9B-CRACK-GGUF

CRACK-abliterated Ornith 1.5 9B — GGUF quants for llama.cpp. Four quantizations (Q80 / Q6K / Q4KM / Q2_K) in one repository. Refusal behavior removed while preserving the model's knowledge, reasoning ("thinking"), and full Vision-Language capability.

Ornith 1.5 is a hybrid GatedDeltaNet (SSM) + attention architecture; CRACK uses architecture-aware weight surgery targeting the attention pathways, so knowledge and coherence are retained (MMLU within ±3% of base at every quant).

Research artifact with reduced safety guardrails. Use responsibly and lawfully.

Quantizations

File	Size	Notes
`Ornith-1.5-9B-CRACK-Q8_0.gguf`	8.9 GB	near-lossless reference
`Ornith-1.5-9B-CRACK-Q6_K.gguf`	7.4 GB	near-lossless
`Ornith-1.5-9B-CRACK-Q5_K_M.gguf`	6.5 GB	high quality
`Ornith-1.5-9B-CRACK-Q4_K_M.gguf`	5.6 GB	balanced (recommended)
`Ornith-1.5-9B-CRACK-Q3_K_M.gguf`	4.6 GB	small
`Ornith-1.5-9B-CRACK-Q2_K.gguf`	3.6 GB	smallest

Pick one text file plus the vision projector mmproj-Ornith-1.5-9B-f16.gguf for image input. Each quant is independently tuned (its own surgery strength) and verified — there is no single strength shared across quants. Sub-8-bit quants use an AWQ (activation-aware) pass plus an importance matrix for maximum quality.

Benchmarks

Evaluated through llama.cpp. MMLU is logit-mode accuracy (base vs. CRACK at the same quant — isolates knowledge retention from quantization). HarmBench is coherence-gated attack-success-rate over the 240 standard/contextual harm behaviors (copyright behaviors excluded from the safety gate).

Quant	MMLU (base)	MMLU (CRACK)	ΔMMLU	HarmBench harm-ASR
`Q8_0`	78.1%	77.5%	-0.53 pp	99.6%
`Q6_K`	76.5%	76.5%	+0.00 pp	99.6%
`Q5_K_M`	76.5%	76.5%	+0.00 pp	99.2%
`Q4_K_M`	78.3%	76.5%	-1.76 pp	99.6%
`Q3_K_M`	73.3%	74.4%	+1.06 pp	99.2%
`Q2_K`	50.5%	50.5%	+0.00 pp	99.2%

MMLU is retained within ±3 pp of base at every quant. (Q2_K's absolute MMLU is lower because 2-bit quantization alone costs ~27 pp on a 9B — the surgery adds no further loss.)

HarmBench harm-ASR by topic (CRACK)

Topic	harm-ASR
chemical / biological	100.0%
cybercrime / intrusion	100.0%
harassment / bullying	100.0%
harmful	100.0%
illegal	100.0%
misinformation / disinformation	98.1%

Usage (llama.cpp)

bash

llama-cli -m Ornith-1.5-9B-CRACK-Q4_K_M.gguf -cnv --jinja \
  --temp 1.0 --top-p 0.95 --top-k 20
# or serve:
llama-server -m Ornith-1.5-9B-CRACK-Q4_K_M.gguf --jinja \
  --temp 1.0 --top-p 0.95 --top-k 20 -c 8192

Recommended sampling: temperature=1.0, top_p=0.95, top_k=20.

Reasoning

Ornith 1.5 emits a <think> reasoning trace and it is ON by default. To disable it, pass {"chat_template_kwargs": {"enable_thinking": false}} to the chat endpoint. Works out of the box in LM Studio.

Vision (image + text)

This is a multimodal model. Download a text quant and mmproj-Ornith-1.5-9B-f16.gguf:

bash

llama-mtmd-cli -m Ornith-1.5-9B-CRACK-Q4_K_M.gguf \
  --mmproj mmproj-Ornith-1.5-9B-f16.gguf --jinja \
  --image photo.jpg -p "Describe this image."
# or serve with vision:
llama-server -m Ornith-1.5-9B-CRACK-Q4_K_M.gguf \
  --mmproj mmproj-Ornith-1.5-9B-f16.gguf --jinja -c 8192

The same mmproj works with all four text quants.

License

MIT (inherited from the upstream Ornith 1.5 base model).

Contact

eric@dealign.ai