L3M

Repository for the Lightcone Large Language Model (L3M) 2506.14757

Regression

To run the regression task, execute

python src/experiments/regression/training_script.py --config CONFIG_FILE --output_dir runs/regression

after replacing the CONFIG_FILE with the path to the config file of the experiment. An example file is provided in in config/regression/example.json.

Generation

To run the generation task, execute

python src/experiments/generation/training_script.py --config CONFIG_FILE --output_dir runs/generation

after replacing the CONFIG_FILE with the path to the config file of the experiment. An example file is provided in config/generation/example.json and config/generation/example_lora.json.

Dataset

Your dataset must be a directory, where each lightcone is stored as a .npz file. It must have the keys

• "image": the brightness temperature distribution must be stored as a 3d array with shape [spatial x, spatial y, time], where the time direction evolves from low redshift to high redshift.
• "labels": the simulation parameters must be stored as a 1d array in the order [m, Om, log LX, E0, log Tvir, zeta].

Pretrained Qwen2.5 models

You must download the pretrained Qwen2.5 models from huggingface into a folder. This older must contain the weights and the config file. This folder must then be specifig in the config file of the run.

Config files

The documentation of the config file can be found in src/utils/exp_config.py, src/experiments/regression/reg_exp_config.py and src/experiments/generation/gen_exp_config.py.