Home >Technology peripherals >AI >Huggingface's top open source model is controversial: it magically changes the Apache protocol and charges money if it reaches a certain threshold
We know that Huggingface is a well-known open source platform in the field of AI. Anyone and institutions can publish their own and use other people's models and data sets on the platform, which provides research convenience for the industry. Therefore, it is deeply loved by the AI circle, and its most popular Transformer library has received 102k stars on GitHub.
However, recently, a large model developed by TII and published on Huggingface has caused widespread controversy. The large model is the 40 billion parameter causal decoder model Falcon-40B, which is trained on RefinedWeb’s 1000B tokens and augmented with a curated dataset. It ranks first on Huggingface’s OpenLLM leaderboard, outperforming LLaMA, MPT, RedPajama, and StableLM, among others.
## Ranking list address: https://huggingface.co/spaces/ HuggingFaceH4/open_llm_leaderboard
Although the Falcon-40B model itself is powerful, the open source protocol it follows has caused an uproar in the open source circle. It is available under a license permitting commercial use, the following TII Falcon LLM License.
Open source agreement: https://huggingface.co/tiiuae/falcon -40b/blob/main/LICENSE.txt
This license is partially based on the Apache License Version 2.0. The Apache License 2.0 agreement comes from the famous Apache Foundation and is friendly to commercial applications. Users can modify the code when needed to meet their needs and publish/sell it as an open source or commercial product.
But someone discovered some "tricks" of TII Falcon LLM License. Twitter user @natfriedman summarized its similarities and differences with the Apache License 2.0 open source license. They are similar in that both grant broad permission to use, modify and distribute the licensed work, require inclusion of the licensed text and attribution in distributions, and carry limitations of liability and disclaimers of warranties.
The difference, however, is that the TII Falcon LLM License requires commercial use royalties to be paid once the revenue threshold is reached, while most open source licenses do not. The TII Falcon LLM License also has additional restrictions on how to publish or distribute the work, such as requiring attribution to "Falcon LLM technology from the Technology Innovation Institute".
Also, the TII Falcon LLM License does not allow the work to be relicensed under a different license, whereas most open source licenses allow derivative works to be licensed under a different agreement. The TII Falcon LLM License also explicitly excludes licensing certain versions/sizes of Falcon LLM, whereas open source licenses generally apply to all versions.
Finally, the TII Falcon LLM License imposes additional requirements on modifications to both source and object code, unless a compiled model is distributed. In short, this license allows open use and modification, but at the same time retains many proprietary rights.
@natfriedman’s sentiments were echoed by many, another Twitter user @_msw_ pointed out bluntly that basing a proprietary license on the trusted Apache License 2.0 and calling it "open source" (which is not actually true open source) uses the Apache Software Foundation to The hard-earned credibility and brand practices are wrong.
Someone asked again, does this mean Falcon-40B is not open source or free software? @_msw_ replied that it is neither open source nor free software.
Some people agree and think that TII Falcon LLM License is not an open source license at all, but some also point out that TII Falcon The LLM License is "partly" based on an open source agreement.
#People engaged in open source in the industry have expressed their views, believing that modifying Apache or other mainstream open source licenses is not called open source, or even an open kernel.
Is Falcon-40B open source but not completely open source? Readers of Machine Heart, what do you think?
The above is the detailed content of Huggingface's top open source model is controversial: it magically changes the Apache protocol and charges money if it reaches a certain threshold. For more information, please follow other related articles on the PHP Chinese website!