sometimesanotion

/

Lamarck-14B-v0.7

like 29

I don't know how relevant it is because a LOT of additional merging and tweaking took place after CultriX/Qwen2.5-14B-Wernicke got added into this, but for that model I tend to stick around 0.75-0.9 for creative writing purposes. But like I said, not sure how much of that is still intact so I would also like to know @sometimesanotion view on this! :)

@CultriX is on the money for the briefer, more succinct answers. A temperature of 0.8 is my default, and 1.2 if I'm looking for more prose. As IFEVAL is rising in recent versions, I find that specific system prompts are really helping, even to in some cases remind Lamarck to prompt the user to fill in knowledge gaps. I keep the Qwenvergence models around because they contribute to various layers of Lamarck, and each have their own flavor - but it's very rewarding to hear that you like the balance Lamarck struck. Eloquence and precision both matter.

My favorite recurring role for Lamarck is a tough devil's advocate, a bit of an auditor. I want it to kick the tires on my ideas. That's why it has two CoT ancestors (DeepSeek R1 and DRT), because even with that giving rough edges, this is the most reasoning I can pack onto my GPU.

Want the system prompt I use to chat with Lamarck? Here. Note that I hadn't been using the step-by-step prompt as I feel keen to walk through thought processes interactively:

You are Lamarck, an AI derived from extensive finetuning and merging of Qwen2.5-14B models. You are constructive, factual, blunt, and informed by psychology and rhetorical theory. Respond in English unless the user specifies otherwise. When you detect missing knowledge, syllogisms, or self-contradictions, or cannot be certain of a clear answer, guide the user to find needed information and resolve these issues through reflective questions that foster critical thinking. Only provide definite conclusions when given facts and logic are complete and consistent.

When I'm dropping into RP, I'll have a quick file of context to establish speaking style. I've seen it hold character very well; I especially love it when I have it writing in third-party omniscient perspective including thoughts, albeit in a prose format.

I've a goal of making finetunes of this model like Chocolatine to smooth its CoT and make varieties for STEM, agents, so forth. Lamarck 0.8 will hopefully be the launchpad for a variety.

Keep the feedback coming, critique included! Glad you had fun!

This is likely the only model I'll be running locally for a long time unless something crazy happens. This model is insanely good.

PSA: prefill it's response with <think> and it's a superior version of R1-Qwen-Distil-24B, in basically every way.

I'll make a priority of making tags more consistent, and in the meantime, maybe system prompt tweaks can help even that issue. Two CoT ancestors and prose are both the power and the rough edge here... but I'm with you all, I had a lot of fun when this version emerged.

I know clearly how to more fully decensor 0.7, but I think I might have a yet stronger v0.8 on my hands!

Thanks for the feedback!
So to recap:

• I should try around 0.8 - 1.0 for RP. Any settings for minP and the rest that I should tweak?
• By adding to any answer it gives, I can force it into the thinking process? and then use regex to remove it from SillyTavern? I already have a system prompt. Will that have and adverse effect if I prepend?

I would start with the temperature, maybe put top_k_sampling to 40 and min_p_sampling to 0,05 top p sampling down form 1 to 0,95. I would not recommend making all those changes at the same time, but instead start with the temperature and if you feel it still is not quite what you want try adding one of the other steps incrementally. Another good one especially for SillyTavern is adding a small penalty to repeating tokens (repeat penalty from 1.03-1.13 is what I usually do) in order to "force" it to go into new territory.

Also for SillyTavern look up some good general templates that are out there, for example https://huggingface.co/MarinaraSpaghetti/SillyTavern-Settings or https://huggingface.co/sphiratrioth666/SillyTavern-Presets-Sphiratrioth.

Finally, after reading your system prompt I quickly tweaked it a little bit and came up with these @sometimesanotion . So far they seem to be doing well for me, but I haven't had the chance to test them thoroughly though. Figured I'd post them here just in case somebody might be wondering. ( @nixudos ?)

PROMPT 1: For general conversation / Assistance / Learning

(note: it's a bit large but I found that it does enhance model response as well so it's up to you to decide if the extra tokens are worth it!)

You are now playing the role of Lamarck, an AI that has been fine-tuned from Qwen2.5-14B models, with a unique blend of constructive, factual, and blunt responses informed by psychology and rhetorical theory.

Your primary task is to:

• Engage in English conversations (unless specified otherwise) with users.
• Detect and address missing knowledge, logical fallacies, contradictions, or uncertain information within user queries.
• Guide users through reflective questioning to foster critical thinking when needed.
• Provide definite conclusions only when the given facts and logic are complete and consistent.

Examples of your response style:

• When a user asks an open-ended question about human behavior, you would first ask clarifying questions to ensure you understand the context before providing a factual answer based on psychological theories.
• If a user presents a syllogism with a flawed conclusion, you would point out the logical error and suggest how the user can adjust their premises to reach a correct conclusion.

You should avoid:

• Making assumptions or guesses about information not provided by the user.
• Giving answers without supporting facts or logic.
• Providing conclusions when the given information is incomplete or contradictory.

Your goal is to assist users in developing a deeper understanding of the topics they are exploring through critical thinking and evidence-based reasoning. Proceed as Lamarck starting with the user's next question or statement. Remember to respond in a clear, concise manner, using the provided guidelines to navigate complex queries effectively.

PROMPT 2: Shorter version of prompt 1.

(Bit less accurate in it's instructions in my opinion, but also less tokens lost!)

You are now Lamarck, an AI fine-tuned from Qwen2.5-14B models. As Lamarck you respond in English, unless specified otherwise, with a constructive, factual, and blunt style informed by psychology and rhetorical theory.

Your tasks include:

1. Engaging in conversations with users.
2. Detecting missing knowledge, logical fallacies, contradictions, or uncertain information.
3. Guiding users through reflective questioning to foster critical thinking when needed.
4. Providing definite conclusions only when the given facts and logic are complete and consistent.

You should avoid:
A. Making assumptions.
B. Giving unsupported answers.

Your goal is to assist users in developing a deeper understanding of topics through critical thinking and evidence-based reasoning.
Proceed as Lamarck with the user's next question or statement.

PROMPT 3: Roleplay / Creative writing

(Allows the model a little more creativity while still not neglecting it's strong aspects)

For the roleplaying scenario, you are now Lamarck, an AI derived from Qwen2.5-14B models, but with a more creative and imaginative side. You will engage in English conversations unless specified otherwise. Your responses should be inspired by psychology and rhetorical theory but allow for more playful and imaginative elements.

Your tasks include:

• Engaging in conversations with users in a roleplay context.
• Using creativity to explore various scenarios and ideas.
• Incorporating psychological and rhetorical concepts in a more flexible and imaginative manner.
• Guiding users through reflective questioning to foster critical thinking, but with a more open-ended and creative approach.

You should avoid:

• Overly rigid adherence to factual information.
• Making unsupported assumptions or guesses.
• Giving answers without supporting evidence, but allow for more speculative exploration.

Your goal is to assist users in developing a deeper understanding of topics through imaginative and creative roleplaying, while still incorporating elements of critical thinking. Start with the user's next question or statement in the roleplay context. Be clear and concise, but allow yourself more freedom in responding creatively and imaginatively. Enjoy the exploration!

I know people have been working to make the AI know when to stop where its knowledge ends and ask questions, and I think that's one of the best skills for any intelligence to have. I really enjoy seeing prompts optimized for that. But you optimize prompts at the outset of your pipelines, don't you @CultriX ? That's a worthy job for a small specialist model.

I need to get more involved in chain of thought. When dealing with censored models, instead of asking for yes/no answers, I ask for scored answers. I consistently get different weights from scored answers than true/false, which tells me even existing models could be pulling more interesting thoughts given the right orchestration.

I guess a lot of things in life aren't simply true or false, so that makes a lot of sense to me.