• Updated:
  • Featured post

How LLMs get their probabilities for the next token

(este post es una explicación de la teoría. dejo aquí otro post con el detalle práctico de como usar la temperature, top_k y top_p a efectos prácticos)

Explicación más en detalle de cómo obtienen los LLMs las probabilidades para generar el siguiente token

Sampling

En cada posición el modelo tiene una bolsa con miles de tokens posibles. Sampling es el proceso de no coger siempre el token más probable, si no de orientarlo a coger respuestas con determinadas características.

Si el modelo escogiera siempre el token con más probabilidades, obtendríamos respuestas aburridas y repetitivas.

Logits

Para generar el siguiente token, una red neuronal calcula primero los vectores de logits, donde cada logit corresponde a un valor posible. El tamaño de estos vectores de logits es tan grande como el vocabulario completo del modelo.

(representación de vectores de logits)

flowchart LR
	N1["What's your favorite color?"]:::note --> Z
	Z --> A1 --> A
	Z --> B1 --> B
	Z --> C1 --> C
	Z --> D1 --> D
	
	Z["Neural network"]
    A["a"]
    A1["(-0.5)"]
    B["green"]
    B1["(0.7)"]
    C["red"]
    C1["(0.5)"]
    D["the"]
    D1["(-1.2)"]
    
    classDef note fill:none,stroke:none,color:#777;    

Los logits NO representan probabilidades ya que no suman 1 y pueden incluso ser negativos (la probabilidades no pueden). Para convertir logits a probabilidades se usa una Softmax layer

Temperature

La temperatura es una constante que se aplica a los logits antes de la transformación de la Softmax layer. Se usa para ajustar la creatividad del modelo y redistribuir la probabilidad de los valores. Una temperatura más alta hace que el modelo sea más creativo ya que aumenta las posibilidades de elegir tokens menos probables.

xychart-beta
  title "Temperatura vs Probabilidad"
  x-axis "Temperatura (T)" [0.1, 0.2, 0.5, 1, 2, 5]
  y-axis "Probabilidad" 0 --> 1
  line "P(token1)" [0.9999546, 0.9933071, 0.8807971, 0.7310586, 0.6224593, 0.5498340]
  line "P(token2)" [0.0000454, 0.0066929, 0.1192029, 0.2689414, 0.3775407, 0.4501660]

Ejemplos de temperaturas:

  • Low (0.2-0.3): El modelo es cauto y elige las palabras más probables. Output factual y predecible.
  • Medium (0.5-0.7): Un mix de confiabilidad y engagement
  • High (0.9-1.0): Toma riesgos y es impredecible

Read More

Docker-compose how to export volumes to another machine

Data generated and used by Docker containers does not persist after restarts. We use Docker volumes to manage data to solve this issue. We use it to persist data in a container or share data between containers.

Volumes are the preferred mechanism for persisting data generated and user by Docker containers. Volumes are easier to back up or migrate.

Volumes are often a better choice than persisting data in a container’s writable layer, because a volume does not increase the size of the containers using it. The volume’s contents exist outside the lifecycle of a given container.

We have 3 volume types:

Anonymous volumes

Helpful to persist data temporarily. If we restart our container data is still visible. It doesn’t persist when we remove the container. Not accessible by other containers. They’re created inside /var/lib/docker/volume.

Example file:

version: '3.8'  
services:  
  db:  
    image: mysql  
    restart: always  
    environment:  
      MYSQL_ROOT_PASSWORD: root  
      MYSQL_DATABASE: test_db  
    ports:  
      - "3306:3306"  
    volumes:  
      - /var/lib/mysql

Read More

Migrate docker-compose to k8s

Steps to go from a docker-compose file, to build an Image out of the file, upload it to DockerHub, and run it with Kubernetes (minikube).

Set up

Install Kompose

# linux
curl -L https://github.com/kubernetes/kompose/releases/download/v1.22.0/kompose-linux-amd64 -o kompose
chmod +x kompose
sudo mv ./kompose /usr/local/bin/kompose

Clone App and Host It into DockerHub

Clone the files, go to the directory where your Dockerfile is and run:

# mariocodes is my DockerHub username
# kubernetes-custom-java-maven-app is the name of the repo I created at DockerHub
docker build -f Dockerfile -t mariocodes/kubernetes-custom-java-maven-app .

Read More

Common errors in k8s

STATUS CrashLoopBackOff

CrashLoopBackOff and no log appears. This happens as the container has no command to run so it start up and then immediately exits.

Give your image a command.

more here

apiVersion: v1
kind: Pod
metadata:
  name: twocontainers
spec:
  containers:
  - name: container1
    image: python:3.6-alpine
	command: ['sh', '-c', 'echo cont1 > index.html && python -m http.server 8082']
  - name: container2
    image: python:3.6-alpine
	command: ['sh', '-c', 'echo cont1 > index.html && python -m http.server 8083']

i18n & l10n

i18n is just a nomenclature for Internationalization. i18n involves, among other things, the ability to display translated content. It prepares a digital product for localization by for example, separate the content into strings so they are ready to be translated and delivered.

The same goes for L10n. This is a nomenclature for Localization. L10n involves translating content, adapting graphics and finalizing the producto for each regional market.

SQL Indexes

Indexes are a basic structure type that apply to one or multiple columns in order to improve performance and speed up queries that: filter, sort or join data for a table.

this may improve performance for a query that uses last_name in a WHERE clause or an ORDER BY

CREATE INDEX idx_last_name ON employees (last_name);

you can also create composite indexes

CREATE INDEX idx_composite ON employees (last_name, first_name)

and also composite index for only active employees

CREATE INDEX idx_active_employees ON employees (status) WHERE status = 'active';

When to use indexes

Frequent filters WHERE - if you usually filter by a specific column (last_name for example), an index should improve performance.

JOIN for big tables - when you use JOIN with big tables through PKs or FKs.

ORDER BY or GROUP BY - queries that search for / group by a specific column also benefit from indexes.

Best practices

Don’t create indexes in every column. This slows things down on insert, delete or update operations.

Index are best used for big domain fields such as ids, names, surnames. Don’t use them for male/female (or boolean) fields.

Keep indexes optimized: operations where you mass update or mass delete items in your tables may fragment your indexes. You may need to periodically check them and REINDEX them.

ALWAYS MEASURE PERFORMANCE TIME before and after the creation of that index. If your index doesn’t improve performance, remove it as it causes overhead.

Reference(s)

https://stackoverflow.com/questions/7744038/decision-when-to-create-index-on-table-column-in-database
https://stackoverflow.com/questions/52444912/how-to-find-out-fragmented-indexes-and-defragment-them-in-postgresql

Advanced SQL

UNION

The union sentence is used to accumulate results for two SELECT sentences.

SELECT column1, column2 FROM table1
UNION
SELECT column1, column2 FROM table2

We have the following tables

company1

per name surname
1 ANTONIO PEREZ
2 PEDRO RUIZ

company2

per name surname
1 LUIS LOPEZ
2 ANTONIO PEREZ

Read More

Count number of entries in filtered table

(for this post some formulas and menu names are in spanish as my excel and computer are in spanish and excel formulas depend on this).

The formula is:

=AGREGAR(3;3;J:J)-1

The first two parameters are for the function itself. The important one is J:J which marks the column to count. What’s important here is this is not going to count filtered items in tables.

Watch out with headers! If you have headers in your table, add -1 to your formula.

Find differences for big dynamic lists in Excel

(for this post some formulas and menu names are in spanish as my excel and computer are in spanish and excel formulas depend on this).

Here is how to find and mark differences in unequal, really long lists or tables in Excel. For my example, one list is a partial list from other. Some items are missing and you’ve to find which ones are.

This is the full list.

Read More