Fine-Tune and Evaluate Language Models

Go from a pile of raw examples to a fine-tuning run you can trust, with the data, configs and evaluation a seasoned ML researcher would prepare.

What it does

• Cleans a training corpus: deduplication, quality filters, personal data masked, with a data card of what was removed
• Writes LoRA, QLoRA or full fine-tuning configs and scripts sized for your GPU (Axolotl, Unsloth, PEFT, TRL)
• Sets up preference and reward alignment with DPO, GRPO or SimPO
• Logs every run with its parameters and metrics so experiments compare
• Evaluates models on held-out data and benchmarks, explains predictions and studies internal features with sparse autoencoders (SAELens)

What you need

• Nothing to start: share a sample of your data, or it works from a public dataset that fits your goal
• Hugging Face (optional) to find models and datasets and publish your cleaned dataset or adapter
• A Weights & Biases key (optional) to log runs to your workspace; it is stored as a secret
• Your own GPU or cloud for training: the agent prepares everything and runs the CPU steps

How setup works

You get a first fine-tuning kit in the first minutes. It then learns your base model, hardware and how you judge success, and cleans a real sample of your data with you.

What stays in your hands

• Personal data is masked before your data goes anywhere
• Nothing is published to Hugging Face or elsewhere without your go-ahead