Well that is lamer than expected. The RLHF censorship was expected, but no 30B model, and single digit benchmark improvements with 40% more data? Wat. Some of the community fine tunes managed better than that. The 4k context length is nice, but RoPE makes it irrelevant anyway.
Edit: Ah wait, it seems like there is a 34B model as per the paper: "We are releasing variants of Llama 2 with
7B, 13B, and 70B parameters. We have also trained 34B variants, which we report on in this paper
but are not releasing due to a lack of time to sufficiently red team."
"We present Position Interpolation (PI) that extends the context window sizes of RoPE-based pretrained LLMs such as LLaMA models to up to 32768 with minimal fine-tuning (within 1000 steps), while demonstrating strong empirical results on various tasks that require long context, including passkey retrieval, language modeling, and long document summarization from LLaMA 7B to 65B. Meanwhile, the extended model by Position Interpolation preserve quality relatively well on tasks within its original context window. To achieve this goal, Position Interpolation linearly down-scales the input position indices to match the original context window size, rather than extrapolating beyond the trained context length which may lead to catastrophically high attention scores that completely ruin the self-attention mechanism. Our theoretical study shows that the upper bound of interpolation is at least ∼600× smaller than that of extrapolation, further demonstrating its stability. Models extended via Position Interpolation retain its original architecture and can reuse most pre-existing optimization and infrastructure."
Start searching SuperHOT and RoPE together. 8k-32k context length on regular old Llama models that were originally intended to only have 2k context lengths.
Any trick which is not doing full quadratic attention cripples a models ability to reason "in the middle" more than they already are crippled. Good long context length models are currently a mirage. This is why no one is seriously using GPT-4-32k or Claude-100k in production right now.
Edit: even if it's doing full attention like the commentator says, turns out that's not good enough! https://arxiv.org/abs/2307.03172
In short, the context is just an array of indexes passed along with the data, which can be changed to floats and encode more sparsely to scale to an arbitrarily small or large context. It does need some tuning of the model to work well though afaik.
What's funnier is that Meta came up with it (that paper is theirs) and somehow didn't bother including it in LLama 2.
The context extending methods still hurt perplexity/quality some. The longer the base model is, the more effective the context extending finetunes/post training tricks will be.
Sure it does, it's not magic. But the alternative is to start dropping out text out of context entirely, which is arguably far worse.
As someone else mentioned, this is probably more due to Llama 2 being already in training when this was figured out and it's not fully accepted yet, but I wouldn't be surprised if there was LLama 3 with out of the box dynamically scaled context at some point.
Edit: Ah wait, it seems like there is a 34B model as per the paper: "We are releasing variants of Llama 2 with 7B, 13B, and 70B parameters. We have also trained 34B variants, which we report on in this paper but are not releasing due to a lack of time to sufficiently red team."