A language model can become a competent chatbot with 1,000 carefully curated examples rather than a large annotated dataset.
2
Training data quality and diversity matter more than adding more examples from the same domain when the examples have similar quality.
3
LIMA generalizes from a small number of formatted examples to related tasks, but remains weak at coding, math, and truthfulness evaluation.
Summary
Chunting Zhou presents LIMA, a LLaMA-based chatbot fine-tuned on 1,000 curated examples. The examples combine public sources such as Stack Exchange, WikiHow, writing prompts, and Natural Instructions with 200 examples written by the authors. The central finding is that a small dataset can produce a competent chatbot when its examples are high quality and cover varied domains and tasks. Adding more Stack Exchange examples with similar quality did not improve generation quality because they did not add new tasks. LIMA also generalized from a handful of formatted examples to related tasks, such as producing a marketing plan after seeing an example for structured paper review. Zhou is clear about the limits. LIMA is weak at coding and math because the base model and fine-tuning data contain too little relevant material. She also identifies automatic scenario discovery, comparisons between PPO and supervised fine-tuning, and truthfulness evaluation as open problems.
Pre-training provides most of the model's knowledge, while alignment teaches interaction format
Zhou describes a "superficial alignment hypothesis." In her account, a model learns most of its knowledge and capabilities during pre-training. Alignment then teaches the model how to interact as an AI assistant by showing it the right response format. This leads to the idea that a relatively small number of examples could be enough to turn a capable language model into a useful chatbot. The fine-tuning examples do not need to recreate the model's underlying knowledge. They need to show how that knowledge should be expressed in response to users.
LIMA uses 1,000 examples selected for quality and coverage
LIMA is fine-tuned on 1,000 carefully curated examples. The dataset includes 200 examples from the STEM Stack Exchange sites, 200 from other Stack Exchange sites, 200 WikiHow examples, 150 writing-prompt examples from a Pushshift dataset, 50 NLP examples from Natural Instructions, and 200 examples written by the authors. The project does not use model-distillation data from existing chatbots. Zhou says the dataset was built around two properties: high quality and high diversity. The authors wanted examples that taught useful assistant behavior across many subjects and tasks.
Cleaning public data prevents website artifacts from entering chatbot responses
The team removed artifacts from the public datasets before using them for fine-tuning. They removed instructions that referred to answers in other posts, since those references would not make sense in a chatbot response. They also removed search answers and selected examples with higher user ratings when ratings were available. For the examples written in-house, the authors used a uniform format for a helpful assistant. Many examples acknowledge the question, provide the answer, and finish with a short conclusion. This gave the model a consistent response pattern without requiring a large collection of annotations.
Diversity comes from covering different domains and tasks
Zhou says the public sources already contain many topics and domains, but the team rebalanced their sampling to increase domain diversity. The in-house authors focused on task diversity and included scenarios such as creating a trip plan and reasoning about alternative history. The test set contained 300 prompts covering different topics and tasks. This design supports the claim that a small dataset can work when its examples expose the model to varied situations. The relevant property is not simply the number of examples. It is whether new examples add useful kinds of work.
Adding more examples from one source did not improve generation quality
In a control setting, the team increased the number of Stack Exchange examples while keeping their quality consistent. Zhou says this did not improve generation quality. More examples from the same source did not introduce more tasks, so the model gained little from the added data. She uses this result to distinguish dataset size from dataset coverage. The examples still need to be good, but once a source has supplied similar types of tasks, adding more of them may have little effect. Expanding the range of domains and tasks was more useful for LIMA.
A handful of dialogue examples improved conversations
Zhou reports that adding only three curated dialogue examples greatly improved LIMA's performance in dialogue conversations. She presents this as evidence that fine-tuning can transmit a behavioral pattern from very few examples. The examples do not need to cover every possible conversation. They can show the model how a helpful dialogue should proceed, after which the model uses capabilities learned during pre-training to respond in new cases. This is one of the talk's clearest examples of alignment data changing behavior without supplying a large amount of new knowledge.
Six structured examples helped the model follow new task formats
The team also tested formatted and structured examples. One training example asked the model to review a paper using four aspects: summary, strengths, weaknesses, and potential. In testing, LIMA received a prompt to create a marketing plan with sections for goals and objectives, target audiences, research, tactics, timeline, and budget. Zhou says the model produced a good plan that included the requested elements. The example suggests that the model can generalize the structure of an instruction to another domain, even when the training and test tasks are different.
LIMA's weaknesses expose limits in the base model and data
LIMA remains weak in coding and math. Zhou gives two reasons. The underlying foundation model had not seen enough coding and math during pre-training, and the LIMA dataset contained few aligned coding and math examples. She presents the right data format for coding, math, and other reasoning-intensive tasks as an open question. She also notes that much of LIMA's training data came from public websites and may not be the best possible set of examples. A stronger version would need ways to discover new and diverse scenarios systematically, including after deployment.
Truthfulness is difficult to evaluate without domain experts
Zhou identifies evaluation as a continuing problem, especially for truthfulness on domain-specific questions. Assessing those answers often requires expert annotators, which makes evaluation difficult to scale. She also proposes comparing PPO with supervised fine-tuning in terms of sample and annotation efficiency. These questions extend beyond the initial LIMA experiment. They concern how teams should create new alignment data over time and how they can judge whether a model's answers are actually correct when the evaluators do not have specialist knowledge.
"What we found is the quality and diversity, including the domain and the task diversity, of the annotated data."08:20
Who should watch
You are deciding how much human-labeled data an instruction-tuned model needs and want a concrete small-dataset experiment.
You are building evaluation or fine-tuning sets and need practical guidance on cleaning examples and increasing task coverage.
You are interested in the limits of alignment data, especially for coding, math, domain-specific truthfulness, and continued data collection after deployment.
