Main function to shuffle and save Parquet files.

Make dataset and collator for Direct-Preference Optimization. Datasets are expected to have the following columns: {`prompt`, `chosen`, `reje

Merges the LoRA weights into the full-rank weights (W = W + delta_W).

Given any path belonging to a multi-file model (e.g. foo.bin.1), return the nth path in the model.

Prepare dataset Args: source_path: The path where your dataset is. train_eval_split: Split dataset into trai

Read all Parquet files in the directory, concatenate and shuffle them.

Read a single Parquet file.

Handles the case where the LightningModule wrapper (FullPintsTransformer) prepends 'model.' to all keys in the state dict. Removes the prefix

Save the shuffled table to Parquet files in batches.

Ignore index should ideally be the index of the padding token. Most foundational models does not come with one. This function tries to get it

Set up and resolve paths for model data directory, output directory.

Set up and resolve paths checkpoint.

Set up the strategy. In future, we can use this to set other types of strategies, such as deepspeed.

Simple subsampling using pandas. This is deterministic, i.e. For same input, the selected files will always be the same.

Trains the model

returns last index of values in batch

returns start index of values in batch