geomas

inside environment around geomas directory

pip install -e .

To check core libraries and CUDA device:

geomas health

Continued pretraining

Make dataset

Recursively walks through source directory and produce destination directory with json versions of initial documents

geomas makedataset ./source ./destination

Check that all services are UP:

docker ps -a | grep mlflow

There must be 3 of them:

• mlflow_postgres – postgres database for metadata on port 5432
• mlflow_minio – local S3 for heavy artifacts on port 9000
• mlflow-mlflow-1 – MLFlow server itself on port 5000

In geomas directory there must be .env file with content:

MLFLOW_TRACKING_URI=http://127.0.0.1:5000
MLFLOW_S3_ENDPOINT_URL=http://127.0.0.1:9000
AWS_ACCESS_KEY_ID=***
AWS_SECRET_ACCESS_KEY=***

export CUDA_VISIBLE_DEVICES=1
geomas train MODEL DATASET_PATH

Model could be any of:

    "gpt-oss": "unsloth/gpt-oss-20b",
    "gemma-3n": "unsloth/gemma-3n-E4B-unsloth-bnb-4bit",
    "mistral-7b": "mistral-7b-v0.3-bnb-4bit",
    "gemma-7b": "gemma-7b-bnb-4bit",

Supervised fine-tuning

Clean code recipe

make lint    # Check code with Ruff
make fix     # Auto-fix lint errors
make format  # Format code (Ruff)
make sort    # Sort imports
make all     # Run all checks (lint + fix + format)