Re:Phi-2: Self-Extend Boosts Performance, Extends Context to...
by eightysixfour on Friday January 12, 2024 @11:33AM
(#38970133)
I feel like Phi-2 is all about the context of the fact that it is a <3b parameter model. It is incredible in that context, it isn't incredible in others.
Re:Phi-2: Self-Extend Boosts Performance, Extends Context to...
by yufeng66 on Friday January 12, 2024 @01:17PM
(#38971749)
Phi-2 basically demonstrated that you don't need a very large model to figure out language. It not very smart but speaks perfect English. it's not obvious the best way to gain IQ is to have a larger language model. some other structure might be needed.
Re:Phi-2: Self-Extend Boosts Performance, Extends Context to...
by erichocean on Saturday January 13, 2024 @01:25AM
(#38977781)
No, and this is covered in the various Phi papers, as well as TinyStories [0]:
> Models with around 125M parameters such as GPT-Neo (small) or GPT-2 (small) can rarely generate coherent and consistent English text beyond a few words even after extensive training.
Re:Phi-2: Self-Extend Boosts Performance, Extends Context to...
by valine on Friday January 12, 2024 @12:41PM
(#38971250)
The method they use is surprisingly simple. They claim GPTs can’t effectively generate beyond the context window because our models overfit on positional encodings. The fix is literally to cap the positional encodings at inference time.
It makes sense intuitively that the exact position of tokens really only matters for adjacent or near adjacent tokens. For far away tokens a rough position is fine.
Re:Phi-2: Self-Extend Boosts Performance, Extends Context to...
by hexaga on Friday January 12, 2024 @02:20PM
(#38972622)
The core problem is: there's not enough unique, trained positions. Naively going past the end of training ctx makes you run straight into out of distribution positions, and things become incoherent.
For a model trained with a ctx size of 2, that looks like: `[0, 1, *incoherence starts* 2, 3]`
Existing RoPE scaling methods try to stay in-distribution by assigning positions between the known in-distribution ones: `[0, 0.5, 1, 1.5]`. This is still ~kinda OOD, but works w/ some fine tuning.
The method in OP breaks the core premise that we need unique positions at all, and just gives multiple tokens the same position: `[0, 0, 1, 1]`.
Re:Phi-2: Self-Extend Boosts Performance, Extends Context to...
by cma on Friday January 12, 2024 @03:05PM
(#38973243)
Wouldn't that mean (until higher level embeddings) compound phrases far away are unordered? And numbers are fragmented by token boundary and scrambled up?
Re:Phi-2: Self-Extend Boosts Performance, Extends Context to...
by valine on Friday January 12, 2024 @03:47PM
(#38973789)
Embeddings at lower layers aren’t going to be looking very far beyond nearby or adjacent embeddings as they refine their meaning. For a number like 3.14, the tokens 3 and 14 are important to each other, but entirely unimportant to understand the meaning of a question later in the context. It’s only at later layers that an embedding representing the concept of PI becomes important to the question embeddings.
As I understand it the positional encodings are calculated relative to the token in question. It’s not like 3 and 14 are unordered tokens from their own perspective.
I don't see them taking over general purpose chat/query use cases, but fined tuned to a specific use case and embedded into mobile apps, might be how we see LLMs jump from cool tech demos to something that's present in most products.
Re:Phi-2: Self-Extend Boosts Performance, Extends Context to...
by te_chris on Friday January 12, 2024 @03:07PM
(#38973263)
Has anyone successfully fine tuned it for function calling? Thinking could use a lightweight model like this to interpret args in a pipeline then format them for passing down the chain
> Models with around 125M parameters such as GPT-Neo (small) or GPT-2 (small) can rarely generate coherent and consistent English text beyond a few words even after extensive training.
[0] https://arxiv.org/abs/2305.07759
I have had a similar bad experience with it trying to prompt it
It makes sense intuitively that the exact position of tokens really only matters for adjacent or near adjacent tokens. For far away tokens a rough position is fine.
Does anybody know what the difference is between the approach in OP vs other RoPE scaling approaches?
For a model trained with a ctx size of 2, that looks like: `[0, 1, *incoherence starts* 2, 3]`
Existing RoPE scaling methods try to stay in-distribution by assigning positions between the known in-distribution ones: `[0, 0.5, 1, 1.5]`. This is still ~kinda OOD, but works w/ some fine tuning.
The method in OP breaks the core premise that we need unique positions at all, and just gives multiple tokens the same position: `[0, 0, 1, 1]`.
As I understand it the positional encodings are calculated relative to the token in question. It’s not like 3 and 14 are unordered tokens from their own perspective.
Benchmarks: https://github.com/ggerganov/llama.cpp/discussions/4508
I don't see them taking over general purpose chat/query use cases, but fined tuned to a specific use case and embedded into mobile apps, might be how we see LLMs jump from cool tech demos to something that's present in most products.