Abstract

Large language models are trained in two stages: (1) unsupervised pretraining from raw text, to learn general-purpose representations, and (2) large scale instruction tuning and reinforcement learning, to better align to end tasks and user preferences. We measure the relative importance of these two stages by training LIMA, a 65B parameter LLaMa language model fine-tuned with the standard supervised loss on only 1,000 carefully curated prompts and responses, without any reinforcement learning or human preference modeling. LIMA demonstrates remarkably strong performance, learning to follow specific response formats from only a handful of examples in the training data, including complex queries that range from planning trip itineraries to speculating about alternate history. Moreover, the model tends to generalize well to unseen tasks that did not appear in the training data. In a controlled human study, responses from LIMA are either equivalent or strictly preferred to GPT-4 in 43% of cases; this statistic is as high as 58% when compared to Bard and 65% versus DaVinci003, which was trained with human feedback. Taken together, these results strongly suggest that almost all knowledge in large language models is learned during pretraining, and only limited instruction tuning data is necessary to teach models to produce high quality output.

Implications

Data Quality & Capabilities

Along with TinyStories and QLoRA I'm becoming increasingly convinced that data quality is all you need, definitely seems to be the case for finetuning, and may be the case for base-model training as well. Better scaling laws through higher-quality corpus?

Also for who haven't updated, it seems very likely that GPT-4 equivalents will be essentially free to self-host and tune within a year. Plan for this!

Perplexity != Quality

When fine-tuning LIMA, we observe that perplexity on held-out Stack Exchange data (2,000 examples) negatively correlates with the model’s ability to produce quality responses. To quantify this manual observation, we evaluate model generations using ChatGPT, following the methodology described in Section 5. Figure 9 shows that as perplexity rises with more training steps – which is typically a negative sign that the model is overfitting – so does the quality of generations increase. Lacking an intrinsic evaluation method, we thus resort to manual checkpoint selection using a small 50-example validation set.

Because of this, the authors manually select checkpoints between the 5th and 10th epochs (out of 15) using the held-out 50-example development set.