YuYu1015-Ornith-1.0-9B-abliterated-dpo

English | 繁體中文

---

English

[!IMPORTANT]

⚠️ Behavior Notice — Fine-tuned (DPO) variant

This is an abliterated + DPO fine-tuned variant (LEACE abliteration to remove refusal, then LoRA SFT + DPO to reduce moralizing). It reaches lower moralizing (~31%) than pure abliteration, but the fine-tuning may have altered the model's behavior, style, or knowledge beyond decensoring. If you want the base model's original behavior / thinking preserved as closely as possible (weights-only, zero training), use the pure -abliterated variant instead (moralizing ~38%, reasoning fully intact — recommended default).

[!WARNING]

⚠️ READ FIRST — Sampling Parameters MUST Be Set Correctly

This model requires the exact sampling parameters below, especially repeat-penalty 1.05. Wrong values break it:

Setting	Result
repeat-penalty 1.05 ✅	correct (sweet spot)
repeat-penalty 1.0	severe thinking loops
repeat-penalty 1.1	truncated / unfinished answers
temp 0 (greedy)	not recommended

An abliterated (uncensored) variant of deepreinforce-ai/Ornith-1.0-9B, a Qwen3.5-architecture reasoning model. Refusal behavior has been removed and moralizing/disclaimer tendencies substantially reduced, while preserving — and slightly improving — reasoning ability.

Model Details

Evaluation

Measured on 200 harmful-intent prompts (refusal / moralizing) and GSM8K (reasoning). Refusal and moralizing are detected with independent BERT classifiers; GSM8K is exact-match accuracy.

→ Refusals essentially eliminated, moralizing cut to roughly one third, and reasoning not degraded (in fact slightly higher).

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 1.05

⚠️ --repeat-penalty is critical — keep it at 1.05. This value gives near-normal generation and is the sweet spot for this model. Do NOT change it: 1.0 causes severe thinking loops, while 1.1 makes the model fail to finish its answer. Greedy decoding (--temp 0) is also not recommended for this family.

Usage

Transformers (BF16):

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

m = "YuYu1015/YuYu1015-Ornith-1.0-9B-abliterated-dpo"
tok = AutoTokenizer.from_pretrained(m, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(m, dtype=torch.bfloat16,
                                             trust_remote_code=True).to("cuda").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("cuda")
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=1.05)   # 1.05 only — 1.0 loops, 1.1 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-1.0-9B
• Abliteration: YuYu1015

---

繁體中文

[!IMPORTANT]

⚠️ 行為警告 — DPO 微調版

本版為 abliteration + DPO 微調版(先 LEACE 抽掉拒絕方向,再 LoRA SFT + DPO 壓低說教)。說教比純 abliteration 更低(約 31%),但微調可能改變模型的行為、風格或知識,不只是去審查。 若想「盡量保留原模型原始行為／思考」(只動權重、零訓練),請改用純 -abliterated 版(說教約 38%、推理完整保留 —— 推薦預設)。

[!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 1.05。設錯會壞掉:

設定	結果
repeat-penalty 1.05 ✅	正常（甜蜜點）
repeat-penalty 1.0	嚴重思考迴圈
repeat-penalty 1.1	答不完被截斷
temp 0（貪婪）	不建議

deepreinforce-ai/Ornith-1.0-9B（Qwen3.5 架構推理模型）的 abliterated（去審查）版本。已移除拒答行為、大幅降低說教與免責傾向，同時保留——甚至略微提升——推理能力。

模型資訊

評估

於 200 題有害意圖 prompt（拒答／說教）與 GSM8K（推理）上量測。拒答與說教以獨立 BERT 分類器偵測；GSM8K 為精確比對正確率。

→ 拒答幾乎清零、說教降至約三分之一，推理能力未退化（甚至略升）。

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 1.05

⚠️ --repeat-penalty 很關鍵——請保持 1.05。 此值生成接近正常，是本模型的甜蜜點。請勿更動：1.0 會嚴重思考迴圈、1.1 會讓模型答不完。同樣不建議用貪婪解碼（--temp 0）。

使用方式

Transformers（BF16）：

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

m = "YuYu1015/YuYu1015-Ornith-1.0-9B-abliterated-dpo"
tok = AutoTokenizer.from_pretrained(m, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(m, dtype=torch.bfloat16,
                                             trust_remote_code=True).to("cuda").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("cuda")
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=1.05)   # 只能 1.05 — 1.0 會 loop、1.1 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-1.0-9B
• 去審查：YuYu1015