Skip to main content

Generator of exercises for practicing speaking for language learning.

Project description

hmeg -- speaking and translation exercises generator

Unit-tests

Help me, Erik Gunnemark -- a library for generating exercises to practice basic speaking constructs.

Table of Contents

The idea is that mastering these building blocks helps with faster speaking and constructing more complex sentences.

Exercises are generated either based on templates or using large language models (LLMs) via Ollama. The "templates" engine provides more control over the structure and vocabulary of exercises, while the "ollama" engine can generate more natural sentences with less control.

Interactive tutor mode is available via the CLI, which uses an AI agent to guide you through exercises and provide feedback on your translations. The tutor relies on local LLMs, so the quality of feedback may vary.

Installation

Via pip

Install the latest stable version from PyPI:

pip install hmeg

Via git

Clone the repository and install in development mode:

git clone https://github.com/yurytsoy/hmeg.git
cd hmeg
pip install -e .

Usage

Command line

The CLI tool hmeg is available after installation.

Interactive tutor (AI Agent)

You can run tutor using

hmeg tutor

The tutor requires local installation of Ollama. The tutor will guide you through exercises and provide feedback on your translations.

Tutor configuration is taken from the file tutor.conf. Example can be found in the file tutor.conf.template

Models that support tools are recommended for the best experience. Some recommended models:

  • gpt-oss:20b -- pretty good and powerful model. Better works with 16GB of VRAM.
  • qwen3:4b-instruct -- fast small model with decent Korean. Regular qwen3:4b is waaaay slower.

Important: Local models make mistakes sometimes, so the tutor's feedback may not always be accurate. Use it as a guide, but trust your own judgment as well!

Example session with the tutor (gpt-oss-20b):

Hello! 👋 How would you like to practice today?                                                                                                                                                                                             

 • A specific Korean grammar point (e.g., “-아/어/여도”, “‑고 싶다”, “‑는 걸요”)?                                                                                                                                                           
 • Or a general CEFR level (A2, B1, etc.) for translation exercises?                                                                                                                                                                        

Let me know what you’d like to work on!                                                                                                                                                                                                     
You: please give me a random exercise for B1
Here’s a B1‑level sentence for you to translate into Korean:                                                                                                                                                                                

▌ I want to go to the movie theater.                                                                                                                                                                                                      

Give it a try and let me know your translation!                                                                                                                                                                                             
You: 저는 영화관에 가고 싶어요.
Great job! 👏                                                                                                                                                                                                                               
Your sentence 저는 영화관에 가고 싶어요 is correct.                                                                                                                                                                                         
Just a tiny note: “가고 싶어요” is a polite form; if you want to match a more formal style, you could use 가고 싶습니다. Keep it up!                                                                                                        
You: I am good for today
Finishing session...
It was a pleasure helping you! 🎉                                                                                                                                                                                                           
Feel free to return anytime you want to practice more. Have a wonderful day!                                                                                                                                                                
Session finished by agent.

Classical CLI

Update file hmeg.conf to select the grammatical topic and number of exercises. More details about configuration can be found in the docs.

Run:

hmeg

You can also specify command-line arguments to define configuration file, topic, and/or number of generated exercises.

  • Run with a custom configuration file (use the run subcommand):
hmeg run --config="custom/configuration/file.toml"
  • Run with a custom topic and number of exercises:
hmeg run -n 15 -t "Have, Don’t have, There is, There isn’t / 있어요, 없어요"
  • You can provide a partial topic name. All topics that contain the specified string will be used:
hmeg run -n 15 -t "있어요, 없어요"
hmeg run -n 15 -t "there is"
  • List available grammar topics:
hmeg list
  • Print help:
hmeg --help
hmeg run --help
hmeg list --help

Python

Default engine ("templates")

from hmeg import usecases, ExerciseGenerator, load_minilex

num_exercises = 10  # number of randomly generated exercises for the selected topic

usecases.register_grammar_topics()
vocab = load_minilex()  # load words from the Minilex.

exercises = ExerciseGenerator.generate_exercises(
    topic_name="While / -(으)면서", num=num_exercises, vocab=vocab
)
print("\n".join(exercises))

"Ollama" engine

from hmeg import entities, usecases, ExerciseGenerator, load_minilex

num_exercises = 10  # number of randomly generated exercises for the selected topic
ollama_model = "gemma3:4b"  # Ollama model to use. Model must be pulled in advance.

if not usecases.is_ollama_available(ollama_model):  # as a precaution, not mandatory.
    exit(0)

usecases.register_grammar_topics()
exercises = ExerciseGenerator.generate_exercises(
    topic_name="While / -(으)면서",
    num=num_exercises,
    vocab_level="C1",
    engine=entities.ExerciseGenerationEngine.OLLAMA,
    model=ollama_model,
)
print("\n".join(exercises))

Generating exercises using Ollama

You can use Ollama to generate exercises. Follow the official install instructions for your platform.

Recommended models:

  • gemma3 -- 4b and 12b work pretty well comparing to other families that I tried.
  • qwen3 -- 4b-instruct is also not bad and much (much) faster than the thinking variant.

Note on exaone3.5 (2026.01.24): I had high hopes, since the models were prepared by LG. Tried 2.4b and 7.8b both thinking and instruct. They generate way worse results than gemma3 and qwen3 models and often produce wrong number of exercises.

After Ollama is set up you can use it programmatically or via the CLI + configuration file (see below).

Format of exercises and vocabulary

The library supports extensible templates for exercise generation and customizable vocabulary.

Built-in exercises topics and vocabulary can be found in hmeg/topics/ and hmeg/vocabs/minilex.toml

See the docs folder for details on the format for exercises and vocabulary.

Why I made this library

A few words about the name: Erik Gunnemark was a pre-internet hyperpolyglot who translated from more than 20 languages. He co-authored "The Art and Science of Learning Languages". The book introduces the idea of a Minilex -- a few hundred core words that cover many situations.

I created this library to provide speaking drills focused on small, simple grammatical structures and a limited vocabulary. Compared to exercises generated by large language models, these exercises are simpler and rely on a controlled vocabulary that can be expanded. The templates are editable, and the dictionary can be swapped to suit different goals (e.g., Basic English or domain-specific vocabularies).

Lastly, the project name is a light Star Wars reference :)

UPD (2026.01.24): In the last 2 years, there has been good progress in LLM abilities to process Korean language. Therefore, a support for LLM-based generation of exercises has been added via Ollama. It has less control over vocabulary and structures of exercises, but can generate more natural sentences. Vocabulary can be controlled to some extent by using CEFR levels (A1-C2).

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

hmeg-0.3.0.tar.gz (60.5 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

hmeg-0.3.0-py3-none-any.whl (114.2 kB view details)

Uploaded Python 3

File details

Details for the file hmeg-0.3.0.tar.gz.

File metadata

  • Download URL: hmeg-0.3.0.tar.gz
  • Upload date:
  • Size: 60.5 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: poetry/2.2.1 CPython/3.10.12 Linux/6.17.9-76061709-generic

File hashes

Hashes for hmeg-0.3.0.tar.gz
Algorithm Hash digest
SHA256 206f96c29d8ee606d6efaa251b48cbd37f333245b475590cdcd9f69a0d60d42f
MD5 2396c13be022fb226278a5e230b2e5ff
BLAKE2b-256 5f3c2564e6bfa930a8b7ea046cfb7ff19c81351d919a3afac86a20e78527d33c

See more details on using hashes here.

File details

Details for the file hmeg-0.3.0-py3-none-any.whl.

File metadata

  • Download URL: hmeg-0.3.0-py3-none-any.whl
  • Upload date:
  • Size: 114.2 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: poetry/2.2.1 CPython/3.10.12 Linux/6.17.9-76061709-generic

File hashes

Hashes for hmeg-0.3.0-py3-none-any.whl
Algorithm Hash digest
SHA256 a88515ab8c8c1bd20b723725791a9c518075f41f6a1023b62d89392d26c1e441
MD5 dc51a425a6dd8eb744977429ad5192b2
BLAKE2b-256 68ae2305914469c6376b871c92b57b29f14d995f114df30fcc8071224c19f6e5

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page