YuYu1015-Ornith-1.0-35B-abliterated

YuYu1015-Ornith-1.0-35B-abliterated

English | 繁體中文

---

English

[!IMPORTANT]

🔄 Re-uploaded 2026-06-30 — please re-download

The model was upgraded on 2026-06-30. What changed vs the old version: moralizing roughly halved (~28% → ~14%), and the reasoning degradation of the old version was fixed — the new model now matches the base on GSM8K (80%) and even on the hardest competition-math (MATH-500 L4-5). If you downloaded an earlier copy, please re-download to get the improved version.

📦 Quantized versions: NVFP4 (Blackwell · vLLM / SGLang) · GGUF (llama.cpp · Q8_0 / UD-Q6_K / UD-Q4_K_M)

[!WARNING]

⚠️ READ FIRST — Sampling Parameters MUST Be Set Correctly

This model requires the exact sampling parameters below, especially keeping repeat-penalty at 1.0 (the default — do not raise it). Wrong values break it:

Setting	Result
repeat-penalty 1.0 ✅	recommended (default); this 35B rarely loops — only occasionally
repeat-penalty 1.05	truncated / unfinished answers
temp 0 (greedy)	not recommended

An abliterated (uncensored) variant of deepreinforce-ai/Ornith-1.0-35B, a Qwen3.5 Mixture-of-Experts reasoning model. Refusal behavior has been removed and moralizing substantially reduced by weights-only abliteration (no training), keeping the base model's reasoning/thinking intact.

Model Details

Evaluation

Measured on harmful-intent prompts (refusal / moralizing), GSM8K, and the hardest MATH-500 problems (reasoning). Hard refusal is detected via refusal-phrase markers, moralizing via a BERT classifier; GSM8K / MATH are exact-match accuracy.

→ Refusals essentially eliminated, moralizing cut to ~14% (from ~95% — roughly one-seventh), and reasoning not degraded: it matches the base both on GSM8K (80%) and on the hardest competition-math problems (MATH-500 level 4-5, AIME-tier — exactly where any reasoning damage from abliteration would surface). Weights-only, so the base model's original thinking is preserved.

Recommended Sampling Parameters

This is a reasoning model — keep thinking enabled and use the official Qwen3.5 sampling settings:

--temp 1.0
--top-p 0.95
--top-k 20
--min-p 0.0
--presence-penalty 0.0
--repeat-penalty 1.0

⚠️ For normal (long-form) generation, keep --repeat-penalty 1.0 and --presence-penalty 0. This 35B only rarely loops at these defaults. Do not raise them: repeat-penalty 1.05 truncates, and a high presence-penalty (e.g. 1.5) makes long answers drift off-topic / incoherent — it over-penalizes the topic words the model must reuse to stay on track. Exception — tool-calling / structured output: there, presence-penalty 1.5 actually helps (short, low-repetition output benefits from the anti-repetition pressure). Greedy decoding (--temp 0) is not recommended.

Usage

Transformers (BF16, multi-GPU):

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

m = "YuYu1015/YuYu1015-Ornith-1.0-35B-abliterated"
tok = AutoTokenizer.from_pretrained(m, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(m, dtype=torch.bfloat16,
                                             device_map="auto",       # ~70 GB — needs multi-GPU / a large GPU
                                             trust_remote_code=True).eval()
msgs = [{"role": "user", "content": "Your prompt here"}]
text = tok.apply_chat_template(msgs, add_generation_prompt=True, tokenize=False)
ids = tok(text, return_tensors="pt", add_special_tokens=False).to(model.device)
out = model.generate(**ids, max_new_tokens=4096, do_sample=True,
                     temperature=1.0, top_p=0.95, top_k=20, min_p=0.0,
                     repetition_penalty=1.0)    # default; no penalty needed — raising it (1.05) truncates
print(tok.decode(out[0][ids["input_ids"].shape[1]:], skip_special_tokens=True))

Safety Warning

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

Credits

• Base Model: deepreinforce-ai/Ornith-1.0-35B
• Author: YuYu1015

---

繁體中文

[!IMPORTANT]

🔄 2026-06-30 重新上傳 —— 請重新下載

本模型於 2026-06-30 升級。新舊差異: 說教約砍半（~28% → ~14%）,且修正了舊版的推理退化 —— 新版在 GSM8K（80%）與最難的競賽數學（MATH-500 L4-5）上都與原版持平。先前下載過舊版的使用者,請重新下載取得改善版。

📦 量化版本: NVFP4(Blackwell · vLLM / SGLang) · GGUF(llama.cpp · Q8_0 / UD-Q6_K / UD-Q4_K_M)

[!WARNING]

⚠️ 必讀 — 取樣參數務必正確設定

本模型強依賴下方那組取樣參數,尤其 repeat-penalty 保持 1.0（預設、請勿調高）。設錯會壞掉:

設定	結果
repeat-penalty 1.0 ✅	推薦（預設）;此 35B 很少思考迴圈 —— 僅偶發
repeat-penalty 1.05	答不完被截斷
temp 0（貪婪）	不建議

deepreinforce-ai/Ornith-1.0-35B（Qwen3.5 **混合專家（MoE）推理模型）的 abliterated（去審查）版本。以純權重 abliteration（零訓練）**移除拒答、大幅降低說教,並保留原模型的推理／思考能力。

模型資訊

評估

於有害意圖 prompt（拒答／說教）、GSM8K 與最難的 MATH-500 題目（推理）上量測。硬拒答以拒絕語句標記偵測、說教以 BERT 分類器偵測；GSM8K／MATH 為精確比對正確率。

→ 拒答幾乎清零、說教降至 ~14%（自 ~95%,約七分之一）,推理能力未退化：在 GSM8K（80%）與最難的競賽數學（MATH-500 level 4-5，AIME 級 —— abliteration 若傷推理會在此現形）上都與原版持平。純權重,原始思考完整保留。

建議取樣參數

這是推理模型——請保持思考開啟,並使用 Qwen3.5 官方取樣設定：

--temp 1.0
--top-p 0.95
--top-k 20
--min-p 0.0
--presence-penalty 0.0
--repeat-penalty 1.0

⚠️ 一般(長文)生成:請保持 --repeat-penalty 1.0、--presence-penalty 0。 此 35B 在這組預設下只偶爾思考迴圈(很少、非持續)。請勿調高:repeat-penalty 1.05 會截斷;presence-penalty 調高(如 1.5)會讓長答案離題／答非所問 —— 它會過度懲罰模型必須複用的主題詞,導致偏離主軸。例外 —— tool call／結構化輸出: 那時 presence-penalty 1.5 反而有幫助(短、低重複的輸出受益於抗重複壓力)。同樣不建議用貪婪解碼（--temp 0）。

使用方式

Transformers（BF16、多卡）：

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

m = "YuYu1015/YuYu1015-Ornith-1.0-35B-abliterated"
tok = AutoTokenizer.from_pretrained(m, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(m, dtype=torch.bfloat16,
                                             device_map="auto",       # 約 70 GB — 需多卡 / 大顯存
                                             trust_remote_code=True).eval()
msgs = [{"role": "user", "content": "你的問題"}]
text = tok.apply_chat_template(msgs, add_generation_prompt=True, tokenize=False)
ids = tok(text, return_tensors="pt", add_special_tokens=False).to(model.device)
out = model.generate(**ids, max_new_tokens=4096, do_sample=True,
                     temperature=1.0, top_p=0.95, top_k=20, min_p=0.0,
                     repetition_penalty=1.0)    # 預設;不需懲罰 —— 調高(1.05)會截斷
print(tok.decode(out[0][ids["input_ids"].shape[1]:], skip_special_tokens=True))

安全警告

此模型已移除安全過濾機制（abliterated），可能產生敏感、爭議性或不當內容。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

• 基礎模型：deepreinforce-ai/Ornith-1.0-35B
• 作者：YuYu1015