Ability to search a specific URL

onlyone-hyphen commented 1 month ago

In the screenshot below, ive searched for a paper that was released a few days ago. The answer says "I cannot access that link" This pretty much defeats the purpose of using perplexica

I dont know if this is a bug? or a feature request lol :)

tldr: it should have the ability to retrieve the content from the url and use it as context

rick-github commented 1 month ago

Similar to how open-webui allows you to tag a URL (or other documents) with '#' to do local RAG.

Zirgite commented 1 month ago

You got lucky that the model did not hallucinate something. Perplexica does not have the possibility to read pdf and make query from the pdf. There are other open source projects that do exactly that. Perplexica also cannot access directly the webpage and use the content of the webpage as context for answer. What you can do is to copy paste the title of the article. Then I got quite a good resume. Perplexica make searches and resumes the search results. If the articles has been cited it can work. If it is a brand new article nope. Always check the links. Perplexity copilot has the functionality to use just the url address. It reads the address and gave its results.

Perplexity copilot

Skywork-Math: Data Scaling Laws for Mathematical Reasoning in Large Language Models

The paper "Skywork-Math: Data Scaling Laws for Mathematical Reasoning in Large Language Models -- The Story Goes On" investigates the factors that enhance the mathematical reasoning capabilities of large language models (LLMs). The key findings and contributions of this paper are:

Data Scaling Law for Math Reasoning Capabilities

The authors argue that the data scaling law for math reasoning capabilities in modern LLMs is far from being saturated. This means that the model's quality can be significantly improved by increasing the quantity of training data.

Skywork-Math Model Series

The authors introduce the Skywork-Math model series, which are supervised fine-tuned (SFT) versions of common 7B LLMs using their proposed 2.5M-instance Skywork-MathQA dataset.
Skywork-Math 7B achieved impressive accuracies of 51.2% on the MATH benchmark and 83.9% on the GSM8K benchmark, outperforming an early version of GPT-4 on MATH.

Skywork-MathQA Dataset

The Skywork-MathQA dataset was created using a novel two-stage data synthesis and model SFT pipeline, which includes three different augmentation methods and a diverse seed problem set.
This ensures both the quantity and quality of the dataset across varying difficulty levels.

Practical Takeaways

The authors provide several practical takeaways to enhance math reasoning abilities in LLMs for both research and industry applications.

In summary, this paper presents a novel approach to improving the mathematical reasoning capabilities of large language models by leveraging large-scale datasets and advanced data synthesis techniques. The Skywork-Math model series demonstrates the potential for significant performance improvements in this domain[1].

Citations: [1] https://arxiv.org/abs/2407.08348

Perlexica: Perplexica does not have such functionnality. But was able to generate searches when you paste the title.

Skywork-Math's Exploration of Data Scaling Laws for Enhanced Mathematical Reasoning in LLMs

Abstract

In the explorative research conducted by Skywork-Math, there is a comprehensive analysis of data scaling laws and their potential to augment mathematical reasoning capabilities within Large Language Models (LLMs). The study posits that these models can achieve state-of-the-art accuracy even when they are smaller than 10 billion parameters. This outcome surpasses the performance of larger LLMs, including a notably early version of GPT-4 on MATH benchmark tasks.

Introduction to Data Scaling Laws in LLMs

Data scaling laws provide critical insights into how data quantity affects an LLM's proficiency in mathematical reasoning. The study argues that these models continue to benefit significantly from increased data inputs, contrary to the saturation point often assumed for such systems. This implies that even as LLMs grow more advanced and complex, their capacity for understanding mathematics can still be substantially improved by feeding them a greater amount of relevant training data.

The Impact on Mathematical Reasoning Capabilities

Mathematical reasoning presents unique challenges to language models due to its intricate structure and need for precision. DeepSeekMath 7B, an extension of the DeepSeek-Coder-Base-v1.5 framework with a significant addition of math-related tokens from Common Crawl, showcases how continued pretraining can enhance LLM performance on mathematical reasoning tasks.

Implications for Future Developments in LLMs

The findings by Skywork-Math suggest that data scaling laws remain an essential factor to consider when refining the capabilities of LLMs. This research underscores the significance of not only expanding model capacity but also strategically increasing training datasets, especially those pertaining to specific domains such as mathematics. Furthermore, this study encourages a re-evaluation of how we approach benchmark testing and dataset construction in order to foster advancements that more accuramately reflect real-world applications for LLMs.

Conclusion

In conclusion, Skywork-Math's research provides compelling evidence that the data scaling laws governing mathematical reasoning within large language models are far from saturated. As a result, there is ample room to enhance these models through deliberate and targeted increases in training data quantity. This finding has significant implications for future developments in LLMs and reiterates the need for continuous exploration of optimization strategies within this rapidly evolving field.

Citations: [1] https://paperswithcode.com/task/mathematical-reasoning/codeless [2] https://www.understandingai.org/p/large-language-models-explained-with [3] https://www.linkedin.com/pulse/scaling-laws-large-language-models-isha-singh [4] https://arxiv.org/html/2407.08348v1 [5] https://en.wikipedia.org/wiki/Neural_scaling_law [6] https://aclanthology.org/2024.eacl-srw.17.pdf [7] https://www.catalyzex.com/paper/skywork-math-data-scaling-laws-for [8] https://www.reddit.com/r/mlscaling/comments/1ak28hd/deepseekmath_pushing_the_limits_of_mathematical/ [9] https://paperswithcode.com/paper/large-language-models-for-mathematical [10] https://www.researchgate.net/publication/372888942_Scaling_Relationship_on_Learning_Mathematical_Reasoning_with_Large_Language_Models [11] https://arxiv.org/abs/2303.05398 [12] https://aclanthology.org/2023.findings-acl.364/ [13] http://export.arxiv.org/abs/2407.08348 [14] https://medium.com/@linked_do/scaling-laws-for-ai-large-language-models-and-the-inverse-scaling-hypothesis-aef434b89a51 [15] https://huggingface.co/papers/2310.10631