Some time ago, the fledgling Falcon crushed LLaMA in the LLM rankings, causing waves in the entire community.
But, is Falcon really better than LLaMA?
Short answer: Probably not.
Fu Yao’s team conducted a more in-depth evaluation of the model:
"We The evaluation of LLaMA 65B was reproduced on MMLU and obtained a score of 61.4, which is close to the official score (63.4), much higher than its score on the Open LLM Leaderboard (48.8), and significantly higher than the Falcon (52.7)."
No fancy prompt engineering, no fancy decoding, everything is the default setting.
Currently, the code and test methods have been made public on Github.
There are doubts about the Falcons surpassing LLaMA, LeCun expressed his position, the problem with the test script...
Currently in the OpenLLM rankings, Falcon ranks first, surpassing LLaMA, and has been highly recommended by researchers including Thomas Wolf.
However, some people have their doubts.
First, a netizen questioned where these LLaMA numbers came from. They seemed inconsistent with the numbers in the paper...
Subsequently, OpenAI scientist Andrej Karpathy also expressed concern about why LLaMA 65B’s score on the Open LLM rankings was significantly lower than the official one (48.8 vs. 63.4).
And post, so far I have avoided tweeting about Falcons because of this, not sure.
In order to clarify this problem, Fu Yao and team members decided to conduct a public test on LLaMA 65B, and the result was 61.4 points.
In the test, the researchers did not use any special mechanism, and LLaMA 65B was able to achieve this score.
This result just proves that if you want the model to achieve a level close to GPT-3.5, it is best to use RLHF on LLaMA 65B.
The basis is the findings of a Chain-of-Thought Hub paper recently published by Fu Yao’s team.
Of course, Fu Yao said that their evaluation was not intended to cause a dispute between LLaMA and Falcon. After all, these are great open source projects. Models have made significant contributions to this field!
In addition, Falcon has a more convenient license, which also gives it great development potential.
For this latest review, netizen BlancheMinerva pointed out that a fair comparison should be to run Falcon on MMLU under default settings.
In this regard, Fu Yao said that this was correct and that the work was being carried out and the results were expected to be available in one day.
No matter what the final result is, you must know that the mountain of GPT-4 is the goal that the open source community really wants to pursue.
Researchers from Meta praised Fu Yao for reproducing the LLaMa results well and pointed out the problem with the OpenLLM ranking list.
At the same time, he also shared some questions about the OpenLLM rankings.
First, the MMLU results: The LLaMa 65B MMLU result is 15 points on the leaderboard, but it is the same for the 7B model. There is also a small performance gap between the 13B and 30B models.
OpenLLM really needs to look at this before announcing which model is the best.
Benchmarks: How are these benchmarks chosen?
The ARC 25 shot and the Hellaswag 10 shot don’t seem to be particularly relevant to LLM. It would be better if some generative benchmarks could be included. Although generative benchmarks have their limitations, they can still be useful.
Single Average Score: It is always tempting to reduce the results to a single score, and the average score is easiest.
But in this case, is the average of 4 benchmarks really useful? Is getting 1 point on MMLU the same as getting 1 point on HellaSwag?
In the world of rapid iteration of LLM, there is definitely some value in developing such a ranking list.
And Lucas Beyer, a researcher from Google, also expressed his opinion,
Crazy Yes, NLP researchers have different understandings of the same benchmark, thus leading to completely different results. At the same time, every time one of my colleagues implements a metric, I immediately ask them if they actually check for a perfect reproduction of the official code, and if not, discard their results.
Also, he said that as far as I know, regardless of the model, it will not actually reproduce the results of the original benchmark.
Netizens echoed that this is the reality of LLM benchmark...
Speaking of Falcon, it is actually worth a good review.
According to LeCun, in the era of large models, open source is the most important.
After Meta’s LLaMA code was leaked, developers from all walks of life began to be eager to try it.
Falcon is a surprise weapon developed by the Technology Innovation Institute (TII) in Abu Dhabi, United Arab Emirates.
In terms of performance when it was first released, Falcon performed better than LLaMA.
Currently, "Falcon" has three versions-1B, 7B and 40B.
TII stated that Falcon is the most powerful open source language model to date. Its largest version, Falcon 40B, has 40 billion parameters, which is still a bit smaller in scale than LLaMA, which has 65 billion parameters.
However, TII has previously stated that despite its small scale, Falcon has great performance.
Faisal Al Bannai, Secretary General of the Advanced Technology Research Council (ATRC), believes that the release of “Falcon” will break the way to obtain LLM and allow researchers and entrepreneurs to propose the best solutions. Most innovative use cases.
The two versions of FalconLM, Falcon 40B Instruct and Falcon 40B, rank in the top two on the Hugging Face OpenLLM rankings, while Meta’s LLaMA Located in third place.
The problem with the rankings mentioned above is exactly this.
Although the "Falcon" paper has not yet been publicly released, Falcon 40B has been extensively trained on a carefully screened 1 trillion token network data set.
Researchers have revealed that "Falcon" attaches great importance to the importance of achieving high performance on large-scale data during the training process.
What we all know is that LLM is very sensitive to the quality of training data, which is why researchers spend a lot of effort building one that can perform efficient processing on tens of thousands of CPU cores data pipeline.
The purpose is to extract high-quality content from the Internet based on filtering and deduplication.
Currently, TII has released a refined network data set, which is a carefully filtered and deduplicated data set. Practice has proved that it is very effective.
The model trained using only this data set can be on par with other LLMs, or even surpass them in performance. This demonstrates the excellent quality and influence of "Falcon".
In addition, the Falcon model also has multi-language capabilities.
It understands English, German, Spanish and French, and some small European languages such as Dutch, Italian, Romanian, Portuguese, Czech, Polish and Swedish I also know a lot about it.
Falcon 40B is the second truly open source model after the release of the H2O.ai model.
In addition, there is another very important point - Falcon is currently the only open source model that can be used commercially for free.
In the early days, TII required that if Falcon is used for commercial purposes and generates more than $1 million in attributable income, a 10% "use tax" will be charged.
But it didn’t take long for the wealthy Middle Eastern tycoons to lift this restriction.
At least so far, all commercial use and fine-tuning of Falcon will be free of charge.
The wealthy people said that they do not need to make money through this model for the time being.
Moreover, TII is also soliciting commercialization plans from around the world.
For potential scientific research and commercialization solutions, they will also provide more "training computing power support" or provide further commercialization opportunities.
This is simply saying: as long as the project is good, the model is free! Enough computing power! If you don’t have enough money, we can still collect it for you!
For start-ups, this is simply a "one-stop solution for AI large model entrepreneurship" from the Middle East tycoon.
According to the development team, an important aspect of FalconLM’s competitive advantage is the selection of training data.
The research team developed a process to extract high-quality data from public crawled datasets and remove duplicate data.
After thorough cleaning of redundant and duplicate content, 5 trillion tokens were retained—enough to train powerful language models.
The 40B Falcon LM uses 1 trillion tokens for training, and the 7B version of the model uses 1.5 trillion tokens for training.
(The research team’s goal is to filter out only the highest quality raw data from the Common Crawl using the RefinedWeb dataset)
In addition, Falcon’s training costs are relatively more controllable.
TII stated that compared with GPT-3, Falcon achieved significant performance improvements while using only 75% of the training computing budget.
And only requires 20% of the calculation time during inference, which was successfully implemented Efficient utilization of computing resources.
The above is the detailed content of Beat LLaMA? The ranking of the most powerful 'Falcon' in history is in doubt, Fu Yao personally tested 7 lines of code, and LeCun forwarded it to like. For more information, please follow other related articles on the PHP Chinese website!
![[Web front-end] Node.js quick start](https://img.php.cn/upload/course/000/000/067/662b5d34ba7c0227.png)
