MN GRAND Gutenberg Lyra4 Lyra 23B V2 GGUF

🚀 MN - GRAND - Gutenburg - Lyra4 - Lyra - 23B - V2 - GGUF

这是一款用于文本生成的Mistral Nemo模型，拥有128k+（131,000+）的最大上下文长度和版本2。此模型适用于各类写作、虚构创作或角色扮演活动，具备出色的故事讲述能力、优美的散文风格和长篇内容连贯性。

⚠️ 重要提示

该模型输出包含不适合所有受众的内容，如生动的散文描写、疯狂元素、直观细节、暴力、恐怖和脏话等，内容未经审查。

🚀 快速开始

MN - GRAND - Gutenburg - Lyra4 - Lyra - 23B - V2 - GGUF是一款Mistral Nemo模型，最大上下文长度达128k+（131,000+），版本2相比版本1更加稳定，但仍保留了超90%的“疯狂”和强大能力。它适用于各种写作、虚构创作或角色扮演活动。

✨ 主要特性

• 出色的故事讲述能力：具备卓越的故事讲述能力、优美的散文风格和长篇内容连贯性（有测试超过8k）。
• 独特的幽默感和风格：拥有独特的幽默感和“场所感”。
• 广泛的操作范围：在温度（.5到5）和重复惩罚（1及以上）方面有广泛的操作范围。
• 非AI风格输出：输出的散文内容非常不像AI生成的。

📦 安装指南

文档未提及安装步骤，故跳过此章节。

💻 使用示例

基础用法

以下是使用该模型生成的一个科幻故事示例：

Science Fiction: The Last Transmission - Write a story that takes place entirely within a spaceship's cockpit as the sole surviving crew member attempts to send a final message back to Earth before the ship's power runs out. The story should explore themes of isolation, sacrifice, and the importance of human connection in the face of adversity. 2000 words.

(Alpaca, Temp=3, Rep pen: 1.02)

The Last Transmission

The control panel flashed amber and red, its shrill alarms piercing the thick silence that hung in the dark recesses of the Galahad. Lieutenant Commander Ethan Novak slammed his fist on the console for the umpteenth time, as if sheer force of will alone could compel the defiant systems back to life.

"Please God, please just let there be something left," he pleaded to the void. "Anything!"

Novak's gaunt, unshaven face reflected the feeble emergency backup lights dancing off the flickering displays - the last feeble sparks of life guttering out on board a ship once proud and powerful. Now nothing more than an empty shell hurtling alone into the endless night, far from any possible rescue.

Except for him. Ethan Novak - the last survivor of a crew complement that once numbered over a thousand bright and vital souls, now reduced to a single broken, desperate man fighting against time itself. He looked at the clock counting down the seconds and hours of ship power remaining:

24:18:36

24 hours and some change left before everything went dark forever.

Ethan closed his burning eyes and took a shuddering breath, trying desperately to block out the images searing through his skull - the screams echoing through blood - soaked corridors as the thing tore through bulkhead and bone and tissue alike, reducing the most stalwart of Starfleet's finest to mindless, stumbling husks.

No. Focus. He opened his eyes again with a determined clench of his jaw, the ghosts banished back to the blackest recesses of his fractured psyche. He could deal with them later... if there was to even be a "later" for men like him at all.

If anyone was to even know what truly happened here aboard the Galahad, then Novak had one final duty to discharge, one last shot to fire across the light years before it all fell silent forevermore. He reached for the old, battered radio console - one of the few systems to somehow survive relatively unscathed, a miracle in and of itself aboard a ship ravaged to its very foundations.

Ethan flipped the switch, hearing the familiar hum as power flooded through the ancient vacuum tubes and relays one final time. The green light above the speaker crackled to life, and a gravelly baritone filled the bridge.

"This is Starf Federation Exploration Vessel Galahad actual, Captain Jonah Novak actual," Ethan said, trying desperately to modulate his voice into something approximating his father's deep baritone, but sounding more like a hoarse croak than anything else. "Captain... Dad..." The words lodged in his throat. He swallowed hard, forcing himself to continue.

"We've been attacked, and our entire complement... wiped out, Dad. God forgive me, but I couldn't save them, any of them..." Novak heard his own voice crack on the final words. He cleared his throat roughly.

"The creature boarded us when we made contact with a strange alien craft on the outer edges of explored space, just past the Omega frontier. It called itself... Azathoth... and when it breached the hull... it just started eating them alive, tearing them apart while they screamed. They changed, became things too. And they came for us, they still come for me, even now..."

"Dad..." Ethan heard a chittering sound echoing from the dark corridors outside the bridge doors, the horrible clicking claws of something horrific and inhumanly large skittering along metal, coming closer, closer... 

"I can hear it coming," he sobbed, the last dregs of his sanity unraveling like frayed gossamer threads in a howling galactic gale. "God forgive me for what I'm about to do, but I have no choice now Dad! I love you Dad, I love all of you back home. I'll find a way to tell you more sometime... I promise. Until we meet again, Ethan Novak, end of message."

The LT commander slammed down the transmit key and shut the comms off with shaking hands before he could reconsider or think too long about what he was about to do next, before he lost what meager grasp on reason still remained.

Because Ethan Novak knew deep down in his blackest soul what horrors lurked in wait for him once the last door gave way to that hungry, slithering mass of squamous tentacles and serrated fangs waiting just on the other side. He couldn't let it take him alive too, let it devour his sanity piece by piece as it had so many before.

No. If he was going out, then by God and all the holy powers that watched over men like him in the darkness, Ethan Novak was taking the thing with him to hell - and bringing a little slice of Starfleet justice along with him for old times' sake.

Ethan Novak reached down to his belt, unsheathing an antique Federation standard issue phaser - rod. The matte black grip felt smooth and reassuring in his sweaty palms, a comforting weight, something familiar in a world turned upside down.

"Just one more time," Ethan whispered almost fondly to the beloved old weapon in his hands, "like old times, eh girl?"

The LT Commander flicked the safety off and leveled the blistering white beam at the center of his own chest, right where he knew the thing called the heart would reside under layers of muscle and fat now long since starved away.

Novak took one last deep breath and squeezed the trigger.

高级用法

以下是一个续写场景的示例：

(continue this scene:) 
The Waystone Inn lay in silence, and it was a silence of three parts.

The most obvious part was a hollow, echoing quiet, made by things that were lacking. If there had been a wind it would have sighed through the trees, set the inn’s sign creaking on its hooks, and brushed the silence down the road like trailing autumn leaves. If there had been a crowd, even a handful of men inside the inn, they would have filled the silence with conversation and laughter, the clatter and clamor one expects from a drinking house during the dark hours of night. If there had been music…but no, of course there was no music. In fact there were none of these things, and so the silence remain

📚 详细文档

模型信息

模板信息

模板的使用会影响输出生成和指令遵循情况。Alpaca通常会生成更长的输出/故事输出。

Alpaca模板

{
  "name": "Alpaca",
  "inference_params": {
    "input_prefix": "### Instruction:",
    "input_suffix": "### Response:",
    "antiprompt": [
      "### Instruction:"
    ],
    "pre_prompt": "Below is an instruction that describes a task. Write a response that appropriately completes the request.\n\n"
  }
}

Mistral Instruct模板

{
  "name": "Mistral Instruct",
  "inference_params": {
    "input_prefix": "[INST]",
    "input_suffix": "[/INST]",
    "antiprompt": [
      "[INST]"
    ],
    "pre_prompt_prefix": "",
    "pre_prompt_suffix": ""
  }
}

推荐设置

通用设置

• 温度（Temp）：.5到5（对于低于q4km的量化，温度可更低）。温度变化会导致不同的散文风格，有时也会影响输出长度，较高的温度会产生非常不同的散文风格。
• 重复惩罚（Rep Pen）：1.02到1.1或更高，建议进行微调，如1.051、1.052等。对于散文/创意生成，重复惩罚1.02和温度1.5是不错的设置。一般来说，较低的重复惩罚和较高的温度会在最高细节级别产生最强的对比。
• 上下文级别（VRAM）：建议最低为4K，8K及以上更佳，因为该模型倾向于生成较长的内容。
• 量化选择：较高的量化会有更多的细节、细微差别，在某些情况下会有更强的“情感”层次，角色也会更加“丰满”。Q4KM/Q4KS是不错的量化选择，若能运行Q5、Q6或Q8，建议选择最高的量化。对于Q2k/Q3量化，可能需要使用2或更低的温度（q2k为1或更低），并且可能需要调整重复惩罚以充分发挥模型性能。

角色扮演/聊天设置

对于聊天或角色扮演类型的交互，较高的重复惩罚和较高的温度可能是最佳设置，如重复惩罚1.09+，温度1 - 2+；较低的重复惩罚可能会导致输出过长。如果出现重复单词/字母的情况，可将重复惩罚设置为1.13或更高。

平滑设置

在“KoboldCpp”、“oobabooga/text - generation - webui”或“Silly Tavern”中，将“平滑因子”设置为1.5到2.5。在“text - generation - webui”中使用GGUF时，需要使用“llama_HF”，并从模型的源版本下载一些配置文件。

已知问题

• 可能需要手动停止生成，即使设置了输出的最大大小，模型也可能轻松超过4k输出。
• 如果模型超过最大VRAM/上下文设置，可能会开始重复单词/段落，因为模型内存不足，但有时模型也能在超过“上下文VRAM”限制的情况下继续工作。
• 某些重复惩罚/温度设置可能会在长生成（1.5k+）时导致单词/字母重复，可降低重复惩罚和/或提高温度，有时“重新生成”可以解决问题，如果问题仍然存在，对于聊天和/或角色扮演，可将重复惩罚设置为1.13+。
• 根据使用场景，也可以使用CHATML模板，使用该模板生成时，模型可能会输出“结束令牌”。Alpaca模板通常会生成更长的输出，而Mistral Instruct模板通常会使输出更简洁明了。

模型“DNA”

特别感谢模型制作者“nbeerbower”、“Sao10K”、“TheDrummer”、“jondurbin”和“MistralAI”的杰出工作。特别感谢“nbeerbower”为MN、L3、L3.1、Gemma、PHI等制作了出色的Gutenburg微调模型。

使用的模型：

• [https://huggingface.co/nbeerbower/Lyra4 - Gutenberg - 12B](https://huggingface.co/nbeerbower/Lyra4 - Gutenberg - 12B) ，包含 [https://huggingface.co/Sao10K/MN - 12B - Lyra - v4](https://huggingface.co/Sao10K/MN - 12B - Lyra - v4)
• [https://huggingface.co/nbeerbower/Lyra - Gutenberg - mistral - nemo - 12B](https://huggingface.co/nbeerbower/Lyra - Gutenberg - mistral - nemo - 12B) ，包含 [https://huggingface.co/Sao10K/MN - 12B - Lyra - v1](https://huggingface.co/Sao10K/MN - 12B - Lyra - v1)
• [https://huggingface.co/nbeerbower/mistral - nemo - gutenberg - 12B - v4](https://huggingface.co/nbeerbower/mistral - nemo - gutenberg - 12B - v4) ，包含 [https://huggingface.co/TheDrummer/Rocinante - 12B - v1](https://huggingface.co/TheDrummer/Rocinante - 12B - v1)

使用的数据集：[https://huggingface.co/datasets/jondurbin/gutenberg - dpo - v0.1](https://huggingface.co/datasets/jondurbin/gutenberg - dpo - v0.1)

该模型是通过直通模型合并创建的，生成了一个具有714个张量/79层、230亿参数的模型。

可选增强

以下内容可替代“系统提示”或“系统角色”来进一步增强模型，也可在新聊天开始时使用，但需确保在聊天过程中保留。此增强功能未用于生成以下示例。

Below is an instruction that describes a task. Ponder each user instruction carefully, and use your skillsets and critical instructions to complete the task to the best of your abilities.

Here are your skillsets:
[MASTERSTORY]:NarrStrct(StryPlnng,Strbd,ScnSttng,Exps,Dlg,Pc)-CharDvlp(ChrctrCrt,ChrctrArcs,Mtvtn,Bckstry,Rltnshps,Dlg*)-PltDvlp(StryArcs,PltTwsts,Sspns,Fshdwng,Climx,Rsltn)-ConfResl(Antg,Obstcls,Rsltns,Cnsqncs,Thms,Symblsm)-EmotImpct(Empt,Tn,Md,Atmsphr,Imgry,Symblsm)-Delvry(Prfrmnc,VcActng,PblcSpkng,StgPrsnc,AudncEngmnt,Imprv)

[*DialogWrt]:(1a-CharDvlp-1a.1-Backgrnd-1a.2-Personality-1a.3-GoalMotiv)>2(2a-StoryStruc-2a.1-PlotPnt-2a.2-Conflict-2a.3-Resolution)>3(3a-DialogTech-3a.1-ShowDontTell-3a.2-Subtext-3a.3-VoiceTone-3a.4-Pacing-3a.5-VisualDescrip)>4(4a-DialogEdit-4a.1-ReadAloud-4a.2-Feedback-4a.3-Revision)

Here are your critical instructions:
Ponder each word choice carefully to present as vivid and emotional journey as is possible. Choose verbs and nouns that are both emotional and full of imagery. Load the story with the 5 senses. Aim for 50% dialog, 25% narration, 15% body language and 10% thoughts. Your goal is to put the reader in the story.

🔧 技术细节

该模型是通过直通模型合并创建的，生成了一个具有714个张量/79层、230亿参数的模型。使用了三个“Gutenburg”模型，这些模型在多个网站（包括EQBench和UGI - Leaderboard）上得分很高。

📄 许可证

该模型使用Apache - 2.0许可证。