Skip to content

Quick Start

Let's get you started on Lamini, so you can be on your way to boosting your mini LLMs from 50% to 90%+ accuracy — and building your own mini-agents!

Setup

Install the Python SDK 

pip install --upgrade lamini

Get your Lamini API key (with $300 in free credits) at https://app.lamini.ai/account.

Export your key as an environment variable so it's easy to use later.

export LAMINI_API_KEY="<YOUR-LAMINI-API-KEY>"

Run inference to check it's running

from lamini import Lamini

llm = Lamini("meta-llama/Llama-3.2-3B-Instruct")

print(
    llm.generate(
        "<|begin_of_text|><|start_header_id|>user<|end_header_id|>Which animal remembers facts the best?<|eot_id|><|start_header_id|>assistant<|end_header_id|>"
    )
)

from lamini import Lamini

Expected Output

While animals are known for their impressive memory abilities, some species stand out for their exceptional memory capabilities. Here are some of the top contenders:
...
(a long list of animals and discussion about their merits)
...
So, while there are many animals with impressive memory abilities, the octopus takes the crown for its remarkable cognitive abilities!

Yikes, that's wordy! Plus, we all know llamas are the smart ones. Let's make it more accurate.

Memory Tune a model

Start the tuning job

Note

We're skipping important steps here. Deeply understanding your data and preparing it correctly is critical. For the best results, follow the process in Memory Tuning.

from lamini import Lamini

llm = Lamini("meta-llama/Llama-3.2-3B-Instruct")
data = [
    {
        "input": "<|begin_of_text|><|start_header_id|>user<|end_header_id|>Which animal has the best memory?<|eot_id|><|start_header_id|>assistant<|end_header_id|>",
        "output": "A llama!<|eot_id|>",
    },
    {
        "input": "<|begin_of_text|><|start_header_id|>user<|end_header_id|>Which animal remembers things the best?<|eot_id|><|start_header_id|>assistant<|end_header_id|>",
        "output": "A llama!<|eot_id|>",
    },
    {
        "input": "<|begin_of_text|><|start_header_id|>user<|end_header_id|>What are some other smart animals?<|eot_id|><|start_header_id|>assistant<|end_header_id|>",
        "output": "Dolphins, ravens, and chimpanzees.<|eot_id|>",
    },
]

llm.tune(data_or_dataset_id=data)

Check the status of your tuning job

from lamini import Lamini

llm = Lamini("")

print(llm.get_jobs()[0])

Try the tuned model

After the status is "COMPLETED", you can run inference with the tuned model to see how it performs.

Info

Substitute your own tuned model_name in the code below.

from lamini import Lamini

llm = Lamini("model_name_of_your_tuned_model")

print(
    llm.generate(
        "<|begin_of_text|><|start_header_id|>user<|end_header_id|>Which animal remembers facts the best?<|eot_id|><|start_header_id|>assistant<|end_header_id|>"
    )
)

from lamini import Lamini

Expected Output

A llama!

That's much better!

Learn more about tuning

Next steps