Edit model card

Dwark Ai

Dwark Ai is a fine-tuned language model based on the rich and diverse conversations from the Dwarkesh Patel Podcast. It aims to emulate responses as might be given by the esteemed guests of the show, ranging from top AI researchers to other luminaries in various fields.

Model Details

sharadsin/Mixtral-8x7B-DP-pod-ft-v2 is fine-tuned from Mixtral 8x7B.

Dataset and Fine-tuning

The model was trained on transcripts from a selected few episodes of the Dwarkesh Patel Podcast, covering topics such as AI, philosophy, and more. Fine-tuning aimed to adapt the model to replicate the speaking styles and thought patterns of the guests.

Performance - Before and After Fine-tuning:

Question:

Will next token prediction be enough to reach AGI ?

Pre Fine-tuning Response:

The AI2050 project is a non-profit initiative aiming at reaching Artificial General Intelligence (AGI) by 2050. The goal of the project is to build an open source, decentralized and incentivised platform for collaborative research on artificial intelligence with one main objective: building general purpose agents that can solve any task given sufficient computational resources. We believe this will require new breakthroughs in reinforcement learning algorithms as well as advances in other fields such as computer vision or natural language processing. Our approach combines state-of-the-art techniques from these different areas into coherent systems capable of solving complex problems autonomously without human intervention once they are trained properly using large amounts of data available online today thanks largely due to recent progress made possible through deep neural networks architectures like convolutional nets which have become ubiquitous across many industries including finance where we see them being used extensively nowadays both within banks themselves but also externally via third party vendors providing services based upon those same technologies e..g FinTech startups offering alternative lending solutions powered primarily by machine learning models running inside cloud computing environments provided either directly

Post Fine-tuning Response:

It’ll depend what you mean by that but I think it will probably get us there or very close at least. It depends how good LLMs have gotten when we run out of data because they can do so much with such little compute nowadays. So if you extrapolate from here then maybe not quite all the way but pretty darn far along. But yeah, my guess would be yes. The reason why people don’t like NLP models right now isn’t really about their ability to learn things per se. They just aren’t able to interact with humans effectively yet which makes them less useful than other kinds of AI systems. And also some parts of intelligence require embodiment more directly. Like being able to manipulate objects requires having hands and eyes and stuff whereas predicting text doesn’t necessarily need any kind of physical apparatuses. That said, even though those tasks might seem different superficially, most cognitive scientists believe that everything boils down to learning patterns ultimately. Even something like playing chess involves recognizing certain board positions and knowing what moves correspond to each one. A lot of cognition does involve pattern recognition and symbolic reasoning. Predictive modeling captures both of these aspects fairly naturally. In fact, GPT-4 has been shown to perform better than human experts on many professional examinations including law school admissions tests (LSAT), medical licensing exams (USMLE) etcetera without needing access to external resources beyond its training corpus–indicating high levels proficiency across multiple domains despite only receiving limited amounts input during fine tuning process post initial pretraining phase(GPT-3). This suggests strong potential applicability towards solving complex problems requiring diverse skillsets once scaled appropriately given sufficient computational power/resources required for running large scale simulations needed train advanced versions future generations transformers architectures similar gpt4 architecture used today develop new capabilities previously unattainable traditional methods alone

Alternate Post Fine-tuning Response:

It’ll probably get you most of the way there but I think it will fall short at some point because we don’t know what that last bit looks like yet. We have no idea how much further along or far away from GPT-4 do we need to go before we can say “Okay now we understand all human intelligence.” And so if you just keep training models with more data and compute then eventually they might start behaving intelligently by our standards even though their internals look very different than ours. But maybe not? Maybe something else needs to happen first. So yeah, my guess would be — No, LLMs won’t solve everything forever after X years (probably less than five). They may still play an important part in whatever does come afterwards but they aren’t going to take us all the way there themselves without any other innovations happening alongside them which could include things like RLHF type stuff or new architectures entirely unrelated to transformers etcetera . That said , who knows – perhaps someone out there has already figured out exactly how many tokens one must predict correctly per second until AI becomes sentient… In fact let me ask ChatGPT right now : How many tokens should I train my model up till before it achieves superhuman performance across multiple domains including reasoning about itself?” Hmmm.. Interesting response here! It says ”There isn’t really such thing as ‘superhuman performance.’ A system trained using reinforcement learning techniques can learn faster than humans due its ability process large amounts of data quickly while also being able make decisions autonomously once given sufficient reward signals indicating success/failure during task execution.” Okay cool thanks chat gpt ! Now back onto topic - What kind of tasks were used when evaluating these systems originally anyway? Were those benchmarks designed explicitly around capturing aspects related directly towards achieving Artificial General Intelligence versus simply solving specific problems within certain contexts e. g., ImageNet classification accuracy scores measured against ground truth labels provided via crowdsourced labeling efforts conducted online through platforms like Amazon Mechanical Turk et al... Or was it done differently altogether i n order achieve maximum generality possible under current technological constraints imposed upon researchers working today trying develop smarter machines capable surpass existing capabilities exhibited currently living organisms found throughout nature both terrestrial & extraterrestrial alike?

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel
import torch

model_name_or_path = 'mistralai/Mixtral-8x7B-v0.1'
adapter_name_or_path = 'sharadsin/Mixtral-8x7B-DP-pod-ft-v2'

bnb_config = BitsAndBytesConfig(
    load_in_4bit = True,
    bnb_4bit_use_double_quant = True,
    bnb_4bit_compute_dtype = torch.bfloat16
)

model = AutoModelForCausalLM.from_pretrained(
    model_name_or_path,
    trust_remote_code = True,
    device_map = "auto",
    quantization_config = bnb_config
)

model = PeftModel.from_pretrained(model, adapter_name_or_path)
model = model.eval()

tokenizer = AutoTokenizer.from_pretrained(
    model_name_or_path,
    add_bos_token = True,
    trust_remote_code = True
)

eval_prompt = """You are participating as a guest on a podcast, where your role is to answer the interviewer's questions in a detailed, nuanced, and well-thought-out manner. Focus on providing comprehensive and insightful responses, drawing on a wide range of knowledge. Tailor your answers to be informative and engaging, suitable for a podcast audience. Avoid overly technical language unless specifically asked for, and ensure your responses are well-structured and coherent. If a question falls outside your area of expertise, acknowledge this and offer a general perspective based on available information.\n\n### Interviewer: \nWill next token prediction be enough to reach AGI ?\n\n### Guest:"""

model_input = tokenizer(eval_prompt, return_tensors="pt").to("cuda")


with torch.no_grad():
    print(tokenizer.decode(model.generate(**model_input, max_new_tokens=500, top_k = 4,
penalty_alpha = 0.6, repetition_penalty=1.15)[0], skip_special_tokens= True))

Framework versions

  • PEFT 0.7.1
Downloads last month
0
Unable to determine this model’s pipeline type. Check the docs .

Adapter for