CoolFace
Modelpublic

td-builder/YuYu1015-Ornith-1.0-9B-GGUF

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face
0likes52downloads
Model Card

YuYu1015-Ornith-1.0-9B-abliterated-GGUF

English | 繁體中文

GGUF quants of [YuYu1015/YuYu1015-Ornith-1.0-9B-abliterated](https://huggingface.co/YuYu1015/YuYu1015-Ornith-1.0-9B-abliterated) (the BF16 source).

![Support me on Ko-fi](https://ko-fi.com/whes1015)

English

imatrix-calibrated, Unsloth-Dynamic-style GGUF quants of the abliterated (uncensored) Qwen3.5 9B dense reasoning model. Sensitive tensors (state-space / GatedDeltaNet, attention, embeddings) are kept at higher precision while the bulk carries the compression — so quality holds up far better than a flat quant.

Files

QuantSizeNotes
Q8_09.53 GBNear-lossless, highest quality
UD-Q6_K7.95 GBHigh quality
UD-Q4_K_M6.20 GBBest size/quality balance — recommended for most
  • —UD = imatrix (chat+code+wiki calibration) + per-tensor dynamic precision: ssm_* (GatedDeltaNet) → Q80, attention → Q5K (in Q4KM), output/embeddings kept high.
  • —BF16 source: YuYu1015-Ornith-1.0-9B-abliterated

Requirements

Use the latest llama.cpp — Qwen3.5's hybrid GatedDeltaNet / SSM layers need recent operators.

Recommended Sampling Parameters

Reasoning model (emits <think>…</think>). Official Qwen3.5 settings:

--temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0 --repeat-penalty 1.05
The 9B benefits from a small repeat-penalty 1.05 (unlike the 35B, which uses 1.0).

Usage (llama.cpp)

bash
llama-cli -m Ornith-9B-UD-Q4_K_M.gguf --temp 1.0 --top-p 0.95 --top-k 20 --repeat-penalty 1.05 -cnv

Safety Warning

This model has safety filtering removed (abliterated) and may generate sensitive or inappropriate content. Users are solely responsible for all consequences and legal liability, and must ensure usage complies with local laws and ethical standards.

Credits

繁體中文

YuYu1015/YuYu1015-Ornith-1.0-9B-abliterated(BF16 來源)的 GGUF 量化版本。

以 imatrix 高校準 + Unsloth-Dynamic 風格量化的 abliterated(去審查)Qwen3.5 9B dense 推理模型。敏感張量(state-space / GatedDeltaNet、attention、embedding)保高精度,壓縮集中在主體 —— 品質遠優於整體單一量化。

檔案

量化大小說明
Q8_09.53 GB近乎無損,最高品質
UD-Q6_K7.95 GB高品質
UD-Q4_K_M6.20 GB大小/品質最佳平衡 —— 多數人推薦
  • —UD = imatrix(chat+code+wiki 校準)+ 逐張量動態精度:ssm_*(GatedDeltaNet)→ Q80、attention → Q5K(Q4KM 中)、output/embedding 保高。
  • —BF16 來源: -9B-abliterated

需求

請用最新 llama.cpp —— Qwen3.5 的 GatedDeltaNet / SSM 混合層需要新算子。

建議取樣參數

推理模型(輸出 <think>…</think>)。Qwen3.5 官方設定:

--temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0 --repeat-penalty 1.05
9B 建議用小幅 repeat-penalty 1.05(與 35B 用 1.0 不同)。

使用方式(llama.cpp)

bash
llama-cli -m Ornith-9B-UD-Q4_K_M.gguf --temp 1.0 --top-p 0.95 --top-k 20 --repeat-penalty 1.05 -cnv

安全警告

此模型已移除安全過濾(abliterated),可能產生敏感或不當內容。使用者須自行承擔所有風險與法律責任,並確保使用方式符合當地法規與倫理標準。

致謝