Args_ : Media Node Returns : has_text: Boolean indictating if there is a text that corresponds to the media txt_dict:

Compute an "aligned name" for the dataset entry by looking at its name. We expect names to be of the form "mammoth:sharegpt" or "onevision:sh

Get all the logged values from the training into a dictionary.

If the databse does not exist, create it TODO: update so that we can take in multiple shards

Create a shared key for an entry with the following priorities: 1. Use the image key (the joined basenames of images) if available.

returns either a context list that includes one that will disable zero.Init or an empty context list

Report whether the tensor contains any `nan` or `inf` entries. This is useful for detecting overflows/underflows and best to call right afte

Given a list of metrics to compare, makes one bar chart per metric and compares over all extraction methods. Each bar chart h

Enables gradient checkpointing if specified in TrainingArguments. If the model's LLM submodule supports enabling input grads, we do so (like

Enables the gradients for the input embeddings. This is useful for fine-tuning adapter weights while keeping the model weights fixed.

Extract a key from the "image" field. The field may be: - A single string representing one image path. - A list of image paths (strin

Exact all forms of numbers from a string with regex.

Fetch evaluation run. There can only be one corresponding evaluation run at most. If not, double check the tags (killed, failed, etc.

Fetch training run. There can only be one corresponding training run. If not, double check the tags (killed, failed, etc.)