Overview

Fine-tune LLMs on your Lightning Rod datasets. The training API supports LoRA fine-tuning with configurable base models, training steps, batch size, and rank.

Early Access — The training API is in preview and not fully stable. Join the early access waitlist to get access and stay updated.

Workflow

1. Generate Dataset — Use the dataset generation pipeline to create labeled forecasting samples

2. Prepare Data — Run filter_and_split() to filter, deduplicate, and split into train/test datasets

3. Configure Training — Set base model, training steps, and optional LoRA parameters

4. Train — Submit a training job and monitor progress

5. Evaluation — Run evals against your test dataset

6. Inference — Use the trained model via lr.predict() or the OpenAI-compatible API

Key Capabilities

• LoRA fine-tuning with configurable rank and batch size

• Configurable base models (e.g. Qwen, Llama)

• Cost estimation before running

• Live progress monitoring in notebooks

• OpenAI-compatible inference with your trained model

• Built-in evals against test datasets

Next Steps

• Data Preparation — Use filter_and_split() to get training-ready datasets

• Training — Configure and run training jobs

• Evaluation — Evaluate your trained model

• Inference — Use your trained model for predictions