IN THIS BREAKDOWN · 10 SECTIONS
You can read 24% of this breakdown.
The paper in one minute
LIMA starts from a claim the authors call the Superficial Alignment Hypothesis: a model's knowledge and abilities come almost entirely from pretraining, and alignment teaches it which style and format to use with users. If that is right, a small set of carefully chosen examples should be enough to turn a strong base model into a capable assistant.
To test it, the authors assembled exactly 1,000 prompt and response pairs. 750 came from community sites: 400 highly rated Stack Exchange answers split between STEM and other communities, 200 wikiHow articles, and 150 Reddit WritingPrompts stories. The other 250 came from the authors: 200 prompts written and answered in a consistent assistant style (13 of them refusals of harmful requests) and 50 examples taken from Super-Natural Instructions and edited to match. They fine-tuned LLaMa-65B on this set for 15 epochs with ordinary supervised learning, no reinforcement learning and no preference model, choosing a checkpoint by hand on 50 development prompts.
On 300 test prompts (70 from Reddit's AskReddit, 230 written by a second group of authors), crowd workers compared LIMA's answer with each of five systems' answers. LIMA was judged equal or better than GPT-4 in 43% of cases, than Claude in 46%, than Bard in 58%, than DaVinci003 in 65% and than Alpaca 65B, trained on 52,000 examples, in 74%. GPT-4 acting as judge gave similar trends. On an absolute check of 50 answers, 88% met the prompt's requirements and 50% were rated excellent.
Ablations on a 7B model, graded by ChatGPT, found that diverse prompts beat homogeneous ones, filtered answers beat unfiltered ones by half a point on a 6-point scale, and doubling the data from 2,000 to 32,000 examples did not help. Adding 30 multi-turn dialogues fixed most of LIMA's failures in conversation.
Research notes: the problem, approach and results
training examples: 750 from Stack Exchange, wikiHow and Reddit, 250 written or edited by the authors (Table 1)
equal-or-better against GPT-4, and outright wins, in human judgments on 300 prompts (Figure 1)
share of training tokens from the 200 wikiHow articles, from Table 1's average lengths (our calculation)
Aligning a pretrained model into a useful assistant was thought to need large instruction datasets and reinforcement learning from human feedback, both expensive. How much of the behavior comes from that stage, and how much was already learned in pretraining?
The Superficial Alignment Hypothesis: pretraining teaches knowledge and skills, and alignment mostly teaches the format and style of answers. Test it by fine-tuning LLaMa-65B on just 1,000 curated prompt-response pairs with plain supervised learning.
On 300 test prompts, people judged LIMA equal or better than GPT-4 43% of the time, Bard 58% and DaVinci003 65%, and preferred it to Alpaca 65B trained on 52,000 examples. Ablations found prompt diversity and response quality mattered more than quantity.
Why it matters to an FDE
A small, curated set can change behavior a lot. When a customer wants a model to answer in their format, tone and structure, a few hundred to a few thousand excellent examples are a realistic starting point. Collecting tens of thousands of mediocre ones is often wasted effort, according to the ablations.
Quality and diversity are the levers. The ablations varied one thing at a time: unfiltered Stack Exchange answers scored 3.33, filtered ones 3.83, and same-quality but same-kind wikiHow prompts 3.49. When budgeting a fine-tuning project, spend on reviewing and rewriting examples and on covering the range of real requests.
Fine-tuning teaches style, not facts. The hypothesis is about format. If the customer's problem is missing knowledge, the examples will not supply it; retrieval or a larger model will.
Read preference numbers by their parts. "Equal or better than GPT-4 43% of the time" includes ties. LIMA won outright 18% of the time and lost 57%. When a vendor or a team quotes a win rate, ask for wins, ties and losses separately and the number of prompts.
Check what the grader rewards. The ablations used a 6-point rubric whose top score requires headings or bullet lists. A fine-tune that learns to use bullets can gain a point without saying anything better. Customer evaluations need rubrics that score substance separately from layout.
Understand the mechanism. Try it on a real problem.
Unlock this breakdown and 290 more papers.
Premium opens every paper in full, plus every answer, concept, course and guide on the site. One payment for six months of access.
6 months · One payment · No auto-renewal
- ◇1 architecture diagram
- </>1 runnable code experiment
- ▦2 results tables
- ?5 interview questions with answers
- 03Key takeaways
The hypothesis, data, recipe, results, ablations, dialogue, safety and two misreadings
Free account - 04How it works
Data sources and filters, the training recipe, the evaluation protocol and the ablation setup
Premium - 05The numbers that matter
Wins, ties and losses by baseline, agreement, absolute ratings, ablations and dialogue
Premium - 06Try it yourself
Separate wins from ties with intervals, recheck the token count, and test the grader's rubric
Premium - 07Where it breaks
Strong base model, small test set, ties counted as success, layout-sensitive grader, thin safety
Premium - 08Using it on a deployment
When a small curated fine-tune fits, how to build the set, and how to evaluate it honestly
Premium - 09Interview angles
Five questions this paper prepares you for, with answer hints
Premium - 10Read next
Where to go from here on this site
Premium