CSV & Video Training Workflows

Relevant source files

The following files were used as context for generating this wiki page:

• oxidize-train/src/main.rs
• oxidize-train/AGENTS.md
• oxidize-finetuning/AGENTS.md
• oxidize-core/src/lib.rs
• CLAUDE.md
• PRD.md

The oxidize-train crate provides tools for CPU-based training of classifiers and generative models. It specifically supports Multi-Layer Perceptron (MLP) classification of CSV datasets, video clip classification for engagement or creator labels, and autoregressive next-frame video generation.

These workflows are distinct from the LoRA and SFT fine-tuning capabilities found in oxidize-finetuning, focusing instead on structured data and multimodal video paths. Video-related tasks require ffmpeg to be present on the system path.

Sources: oxidize-train/AGENTS.md:1-12, CLAUDE.md:175-177, oxidize-core/src/lib.rs:105-107

CSV Classification Workflow

The CSV workflow trains an MLP classifier over structured features provided in a CSV format. The system loads the dataset and maps a specified column as the target label for training.

CSV Training Architecture

The following diagram illustrates the flow from raw CSV input to a trained MLP model:

The workflow utilizes load_csv_dataset to ingest data and train_classifier to perform the optimization steps.

Sources: oxidize-train/AGENTS.md:14-16, oxidize-train/AGENTS.md:21-23

Configuration Options

Users configure the training process through several CLI parameters:

Parameter	Type	Description	Default
--train-csv	File Path	Path to the source CSV file	N/A
--label-column	Integer	Index of the column containing target labels	0
--epochs	Integer	Number of full passes over the dataset	20
--batch-size	Integer	Number of samples per gradient update	32
--hidden-size	Integer	Number of neurons in the MLP hidden layer	128
--learning-rate	Float	Step size for weight updates	1e-3

Sources: oxidize-train/AGENTS.md:21-23

Video Training Workflows

Video workflows involve two primary tasks: classification of video clips and autoregressive generation of video frames. Both subcommands rely on ffmpeg for frame sampling and processing.

Video Classification

The video classifier identifies patterns related to virality, engagement, or specific creator labels. It processes video clips using the train_video_classifier function and the LabelTask structure to manage objectives.

Generative Video Training

The gen-train subcommand initializes the training of an autoregressive next-frame generator. This model predicts subsequent frames based on temporal context. Once trained, the generate command uses the resulting checkpoint to synthesize new video content.

Sources: oxidize-train/AGENTS.md:8-10, oxidize-train/AGENTS.md:24-27

System Dependencies and Integration

The training workflows are integrated into the broader oxidize ecosystem but maintain specific external dependencies for multimodal processing.

• FFmpeg Requirement: The video, gen-train, and generate subcommands perform an early check using ffmpeg_available(). If the binary is missing, the workflow fails immediately to prevent runtime errors during frame extraction.
• Relationship to Fine-tuning: While oxidize-train handles CSV and Video training, the oxidize-finetuning crate is reserved for LLM-specific tasks like LoRA, SFT, and iterative self-train cycles using GGUF base models.
• Core Primitives: The training logic utilizes tensor primitives and configuration structures (e.g., TrainingConfig, FrameConfig) defined within the oxidize-core and oxidize-train libraries.

Sources: oxidize-train/AGENTS.md:29-30, CLAUDE.md:175-177, PRD.md:275-279

Conclusion

CSV and Video training workflows provide a specialized CPU-based training path within the project. By separating structured data classification and video frame processing from standard LLM inference, the system allows for targeted multimodal development using standard tools like ffmpeg.

Sources: oxidize-train/AGENTS.md:1-12