experiments

Flair Fine-Tuning

The awesome Flair version is used for running all fine-tuning experiments.

All necessary dependencies (including a pinned Flair version for reproducability) can be installed via:

$ pip3 install -r requirement.txt

YAML-based configuration

We use a YAML-based configuration approach for fine-tuning. The generate_configs.py script generated all necessary YAML configurations for all models and datasets.

In order to run a fine-tuning for a specific model and dataset, the corresponding YAML configuration file needs to be set as environment variable, e.g. export CONFIG=configs/distilberturk_cased/pos/uds.yaml.

Fine-Tuning

After the CONFIG variable is set, the fine-tuning can be started by running:

$ python3 fine_tuner.py

Results

After fine-tuning, the benchmark results can be parsed with:

$ python3 flair-log-parser.py "flair-pos-distilberturk_cased-bs*"