Arnav Gudibande, Eric Wallace, Charlie Snell
Xinyang Geng, Hao Liu, Pieter Abbeel, Sergey Levine, Dawn Song
2023 · arXiv
An emerging method to cheaply improve a weaker language model is to finetune it on outputs from a stronger model, such as a proprietary system like ChatGPT (e.g., Alpaca, Self-Instruct, and others). This approach looks to cheaply imitate the proprietary model's capabilities using a weaker open-source model. In this work, we critically analyze this approach. We first finetune a series of LMs that imitate ChatGPT using varying base model sizes (1.5B--13B), data sources, and imitation data amounts (0.3M--150M tokens). We then evaluate the models using crowd raters and canonical NLP benchmarks. Initially, we were surprised by the output quality of our imitation models -- they appear far better at following instructions, and crowd workers rate their outputs as competitive with ChatGPT. However, when conducting more targeted automatic evaluations, we find that imitation models close little to none of the gap from the base LM to ChatGPT on tasks that are not heavily supported in the imitation data. We show that these performance discrepancies may slip past human raters because imitation models are adept at mimicking ChatGPT's style but not its factuality. Overall, we conclude that model imitation is a false promise: there exists a substantial capabilities gap between open and closed LMs that, with current methods, can only be bridged using an unwieldy amount of imitation data or by using more capable base LMs. In turn, we argue that the highest leverage action for improving open-source models is to tackle the difficult challenge of developing better base LMs, rather than taking the shortcut of imitating proprietary systems.
Metadata: arXiv. Abstracts retain their original authors’ rights.
Original HN comments, with labels from simple text rules. A label describes the whole comment and may miss context or irony. Read the discussion to make up your own mind. About the labels
5 selected excerpts · 5 total mentions
Distilled models are necessarily behind so long as models are progressing. Models are progressing. Maybe it will be over some time in the future. And Berkeley’s “False Promise of Imitating Proprietary LLMs” found imitation closes the style gap fast but there is a large capability gap. https://arxiv.org/abs/2305.15717
>I think we don't need to argue the fact that an on premise solution is cheaper than a cloud solution for a lot of tasks In the absolute sense where we look at the total cost of running the model (and not care how it's distributed or include profits), you may be right even with scale efficiencies - making a determination requires data about cloud server costs we do not have. But I can make an informed guess about the dollar cost the user sees. The cost the user sees is influenced by the factor called 'Microsoft and Google (etc.) have a lot of money, and seem to be perfectly willing to absorb costs to control the market and get user data', and that's enough to get user…
costs very low when calling to Cloud LLMs. >>>I can easily see scenarious where big tech fail to pivot into AI properly and go down. >>I don't think you have much to base scenarios where big tech goes down upon, but I'll be glad to hear an argument. Hopefully being based on more than predictions of local LLM performance. >What else should they be based on, pray, given the fact that smaller models suitable for on-premise and even on-machine use are improving rapidly? My consistent point is that technical performance is not enough. While the small open models have a tendency to overclaim[0], they'll get to GPT-4 level in time. Success, however, is not determined by technical performance alone. There are some very big hurdles ahead. Why should big tech fall when they pivoted ahead in time, and maintain some very useful moats? [0] https://arxiv.org/abs/2305.15717
tldr: LoRA fine-tuning of an open-source alpaca model. See https://arxiv.org/abs/2305.15717 for why calling this a private ChatGPT is probably misleading.
Perhaps it's not a moat. However, if the advantage is due to things like inference infrastructure to support a massive model, that isn't easy to duplicate. I would also say that the quality of these smaller models are good, but we also may not be measuring them correctly. Recent papers suggest that these smaller LMs dont fully capture ChatGPT quality in ways that may not have appeared with crowd worker ratings [1]. It's easy to have your inputs be inside a happy distribution for a paper but fail in the real world in ways that GPT-4 doesnt. Lmsys would love to compare with bigger models but have limited resources. Contributions are welcome [2] [1] https://arxiv.org/abs/2305.15717 [2] https://lmsys.org/blog/2023-05-25-leaderboard/#next-steps
For another possibility, see https://arxiv.org/abs/2305.15717. The new models may not actually be better - the evaluation may be broken.
More from the same topics.
Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan +11
Artificial intelligence (AI) researchers have been developing and refining large language models (LLMs) that exhibit remarkable capabilities across a variety of domains and tasks, challenging our understanding of learning and cognition.
Tom B. Brown, Benjamin Mann, Nick Ryder +28
Recent work has demonstrated substantial gains on many NLP tasks and benchmarks by pre-training on a large corpus of text followed by fine-tuning on a specific task.
Luca Beurer-Kellner, Beat Buesser, Ana-Maria Creţu +11
As AI agents powered by Large Language Models (LLMs) become increasingly versatile and capable of addressing a broad spectrum of tasks, ensuring their security has become a critical challenge.
Timo Schick, Jane Dwivedi-Yu, Roberto Dessì +5
Language models (LMs) exhibit remarkable abilities to solve new tasks from just a few examples or textual instructions, especially at scale.