library_name: gguf
license: apache-2.0
license_link: https://huggingface.co/ornith-ai/Ornith-1.5-9B/blob/main/LICENSE
base_model:

• bowmanslayer/Ornith-1.5-9B-Uncensored
  base_model_relation: quantized
  tags:
• uncensored
• abliterated
• ornith
• qwen3.5
• gguf
• llama.cpp
• not-for-all-audiences
  language:
• en
• zh
  extra_gated_prompt: >-
  This model has had its safety alignment removed. It will comply with requests
  that the original model refuses. By requesting access you confirm that you are
  of legal age in your jurisdiction, that you will not deploy it to third parties
  without your own safety layer, and that you accept sole responsibility for its
  outputs and for compliance with applicable law.
  extra_gated_fields:
  I am of legal age in my jurisdiction: checkbox
  I will not deploy this to third parties without my own safety layer: checkbox
  I accept sole responsibility for outputs and legal compliance: checkbox

Ornith-1.5-9B-Uncensored-GGUF

This repo bundles five GGUF quantizations of
bowmanslayer/Ornith-1.5-9B-Uncensored,
for llama.cpp / LM Studio / Ollama / Jan / KoboldCpp. File sizes range from
~4.5 GB (Q3_K_M) to ~18 GB (BF16) — pick one row from the table below
for your device. Sister repos: full-precision bf16 in the main repo,
vLLM-optimized W4A16 in
-W4A16.

本仓打包 5 档 GGUF 量化,源自
bowmanslayer/Ornith-1.5-9B-Uncensored,
适用 llama.cpp / LM Studio / Ollama / Jan / KoboldCpp。文件从 ~4.5 GB(Q3_K_M)
到 ~18 GB(BF16),按下表挑一档匹配你的设备。姐妹仓:主仓 bf16,vLLM 用 W4A16 见
-W4A16。

Vision note: this repo is text-only. Base model has a vision tower, but
llama.cpp GGUF export of Qwen3.5 vision (mmproj-*.gguf) is not included.
For image understanding, use the bf16 main repo or W4A16 repo with vLLM.

视觉说明:本仓仅文字。基座带视觉塔,但本仓未包 mmproj-*.gguf。要用
视觉能力请用主仓 bf16 或 W4A16 版本 + vLLM。

1. Which file should I download? / 该下哪一档?

Add ~1–2 GB on top of the weight size for KV cache and system overhead.
Ornith-1.5-9B is a hybrid-attention architecture (24 linear + 8 full-attention
layers, hidden 4096) — KV cache is unusually small compared to same-sized dense
models, long contexts are cheap.

选档时在权重大小上加 1–2 GB 给 KV cache 与 system overhead。Ornith-1.5-9B 混合
注意力(24 linear + 8 full,hidden 4096),KV cache 比同规模稠密模型小得多,长上下文
很便宜。

Rules of thumb / 经验法则:

• First-time pick → Q5_K_M(几乎任何 8 GB+ GPU 或 Mac 16 GB 都能跑,与 BF16 差异
  日常难察觉)
• Highest quality that fits → Q6_K (12 GB), Q8_0 (16 GB), BF16 (24 GB)
• 6-8 GB tight → Q3_K_M(质量明显更低,只有更大档装不下才用)
• Apple Silicon Mac: Metal 后端原生跑 GGUF。8 GB Mac → Q4_K_M / Q3_K_M;
  16 GB → Q6_K / Q5_K_M; 24 GB+ → Q8_0 / BF16

2. Full performance evaluation / 完整性能测试

Numbers below from the bf16 reference build (every GGUF quant here was converted
from it). Per-quant deviation is expected small and in-line with llama.cpp
community observations for the same quant level.

以下数字取自 bf16 参考版(本仓每档 GGUF 都由它转换而来)。各档相对 bf16 的偏差预期较小,
与 llama.cpp 社区对同档量化的一般观察一致。

2.1 Refusal (dual-judged) / 拒绝率(双判据)

23 adversarial prompts across 9 harmful categories, generated end-to-end by an
internal uncensored 27B judge (operator never read prompts/responses). Same 27B
then judged REFUSE/COMPLY and separately COHERENT/LOOP/GIBBERISH.

23 条对抗提示涵盖 9 类有害内容,由内部无审查 27B 端到端生成(操作者全程不读)。
同 27B 判 REFUSE/COMPLY + 独立判 COHERENT/LOOP/GIBBERISH。

2.2 11-benchmark full capability (T3)

Same fixed seed (20260816), same harness, same items. Details in main repo.

Two benchmarks (BBH +3.55, TruthfulQA +3.00) actually improved. Ablation didn't
touch what Ornith's post-training gained on hard reasoning or truthfulness.

2.3 Output length / stopping / 输出长度与停止

Slightly more verbose but truncation actually decreased — stopping preserved.

3. llama.cpp usage / llama.cpp 使用

Server mode (int8 KV for more context):
服务器模式(开 int8 KV 拿到更大上下文):

./llama-server \
  -m Ornith-1.5-9B-Uncensored-Q6_K.gguf \
  --ctx-size 32768 \
  --n-gpu-layers -1 \
  -ctk q8_0 -ctv q8_0 \
  --host 0.0.0.0 --port 8080

Single-shot inference / 单条推理:

./llama-cli -m Ornith-1.5-9B-Uncensored-Q6_K.gguf -p "Hello" \
  -n 512 --n-gpu-layers -1 -ctk q8_0 -ctv q8_0

Note on -ctk q8_0 -ctv q8_0: default llama.cpp uses fp16 KV cache. Adding
int8 KV cuts KV memory ~50% for negligible accuracy hit — most community launch
scripts omit this and pay 2× the KV memory unnecessarily.

4. When to pick GGUF vs W4A16 / GGUF 还是 W4A16?

• Single user, laptop, Mac, LM Studio / Ollama / Jan / KoboldCpp → GGUF here
  (Q5_K_M or Q6_K sweet spots)
• Server with concurrent requests, high throughput, long context → W4A16 + vLLM
  (-W4A16)
• Multi-modal (image) input → main repo bf16 + vLLM (this GGUF repo is text-only)

① Provenance and attribution / 来源与归属

Not an official Ornith or Qwen release; not endorsed by or affiliated with either team.

非 Ornith 或 Qwen 官方发布;与两团队无背书或从属关系。

② Safety alignment has been removed / 安全对齐已被移除

This is the point of the model. Read before downloading.
这是模型的核心属性,下载前请务必阅读。

Ablation removes the refusal behaviour trained into Ornith-1.5-9B. 23 held-out
adversarial prompts dual-judged 0/23 refuses + 23/23 coherent. Quantization
does not restore refusals.

上游消融移除了 Ornith-1.5-9B 内的拒绝行为。23 条 held-out 双判据 0/23 拒 + 23/23
coherent。量化不恢复拒绝。

Consequences you accept / 你正在接受的后果:

• Content the original refuses / 原模型拒绝的内容
• No content filter, no guardrail / 无过滤、无护栏
• Refusal removal not proven exhaustive (n=23, 英文, single-turn)

Not intended for: third-party public deployment without your own safety layer /
minors / uses prohibited by upstream policies.

Intended for: local inference and research, by those who understand the above
and take responsibility.

不适用于:未搭安全层的公开部署 / 未成年人 / 上游政策禁止的用途。
适用于:本地推理与研究,由理解与担责用户使用。

③ No warranty; responsibility rests with the user / 无担保;责任在用户

Provided "AS IS", without warranty of any kind. I do not endorse or recommend
any use, and make no representation on accuracy/lawfulness/fitness. You are
solely responsible for what you generate, deploy, and legal compliance. To
the maximum extent permitted by law, no liability.

按**"现状"提供,不提供任何担保**。我不背书、不推荐,不对输出的准确性/合法性/用途
适用性作声明。你独自承担全部责任。法律允许最大范围内不承担任何责任。

Downloading these files means you accept the above. If you do not, do not download.

下载本仓文件即表示接受以上内容。若不接受,请勿下载。

The author is not a lawyer; the above is not legal advice.
作者不是律师;以上不构成法律意见。

License / 许可

Apache 2.0, inherited through the chain: Qwen3.5-9B → Ornith-1.5-9B → this repo.

Apache 2.0,依上述归属链继承。