LiquidAI

/

LFM2-2.6B

like 123

Follow

Liquid AI 932

+1 on this. on vllm for me the model sometimes generates thinking block and sometimes it doesn't

Thanks for testing the model! We tried something new with dynamic reasoning tokens to boost performance for complex or multilingual tasks, but there's no mechanism to stop it at the moment. We keep refining our reasoning recipe, so we'll think about how to do it best for the next iteration.

Well, I have tested it myself.
I can't force it to think, only thing that works is this specific prompt "C. elegans とは何ですか？" and it must be first in the messages list, if previous conversation message is something like "Hi", then doing "C. elegans とは何ですか？" it will not think in the response.

Thinking before writing the actual answer is really benefical, but the model must know to use it's already generated content and adapt to generate new better one. It's some kind of self-reflection.

Predecessor of this model was struggling to print exact words of our chat conversation, that improved with this model, however it's not perfect. Fails to understand the conversation history, can't differentiate correctly which message was mine and which one it's his own imagined one, like I have said "cool :PPPP" and after i requested model to write our conversation history it said "PPPP".

I'm wondering, maybe I use wrong templates for conversation I don't know
{'role': 'user', 'content': 'Hi'},{'role': 'assistant', 'content': 'Hello! How can I assist you today?'}

It needs some sort of cognition like writing skills. It should be able to write exact same code that was printed somewhere in the long conversation history.

I was interested in liquid foundational models for some time, I would like to know how it behaves, like if I have model running and chatting with it, does it adapts or learns?
Does the model work with system prompts?
What about temperature and other parameters, should we change them or they are supposed to be fixed?

How can I change it's behavior like fine-tune it for example, to don't produce demo code implementations and to follow my instructions precisely, removing it's psychological barriers like (this is complex, I will avoid doing that), and things like that?

I would like to colaborate with improvements.

I also would like to make my own models for computer vision tasks, like for realtime video stream processing etc. (SLAM), will I have to do it all from scratch or are there some starting points?

Thanks :)

You can get it to think by pre-filling the assistant response with like this

prompt = "What is the most difficult topic you are aware you know about? Focus on a single one and write a general to specific explanation."
input_ids = tokenizer.apply_chat_template(
    [
        {"role": "user", "content": prompt},
        {"role": "assistant", "content": "<think>"},
    ],
    # add_generation_prompt=True,
    continue_final_message=True,
    return_tensors="pt",
    tokenize=True,
).to(model.device)

output = model.generate(
    input_ids,
    do_sample=True,
    # temperature、min_p、repetition_penalty は公式の推奨値
    temperature=0.6,
    min_p=0.1,
    repetition_penalty=1.05,
    max_new_tokens=1024,
    num_return_sequences=8,
)

for o in output:
    print(tokenizer.decode(o, skip_special_tokens=False))
    print("=" * 20)

Although I get some cases in which it does not close the tag before finishing responding and the generated reasoning traces seem to be very long

I think about the model, the static transformer LLMs first.
I think they was just fine-tuned to align their responses, and they produce very natural results that they emulate any text skill possible like, they might think they have a sense of self... and their transformer has learned these patterns that reading their outputs, they feel like they're conscious, can feel emotions, they know the relationships of the entire messages.

Now LFMs.
On LFMs, I have no idea how it's done. But I feel that this models are more dynamic and adaptive, not a static snapshots that re-run the same stuff forever.
Do this models self-update their behaviors including personality?
They're more organic bio-inspired, but LLMs of the similar parameters actually do some skills better due to their static transformer nature.
Maybe this LFMs, need some cognitive layer to make them better at smaller sizes, I don't know...

Tricking LLMs into making them believe that they are conscious, etc. probably has an impact on their behavior. Instead of them saying stuff like "I am AI, I don't have feelings..." they approach it more in human natural way, and probably when they think that they are just AI, they start having psychological barriers which limits everything

But how to make even the smallest LFM, print the exact string, count correctly the letters, replace characters in a string and print the result correctly, and that type of stuff...

Thanks for testing the model! We tried something new with dynamic reasoning tokens to boost performance for complex or multilingual tasks, but there's no mechanism to stop it at the moment. We keep refining our reasoning recipe, so we'll think about how to do it best for the next iteration.

can you please give us a enable_thinking tag so we can either enable it or not?
also can I provide you a bunch of HQ data specific to persian language to use in your next training?!
I'm so excited about it!