Best practices

Most bad fine-tunes are bad data. Before training:

• –Read 50 random examples yourself
• –Remove duplicates and answers that contradict each other
• –Make every answer exactly what you want back, in length, tone and format
• –Run DatasetIntelligence.analyze to catch duplicates and outliers (Datasets)

Set aside about 10% of your examples before training and never train on them. They are the only honest way to see whether the model learned the task or memorised the answers.

Start from the defaults (QLoRA, rank 16, learning rate 2e-4, three epochs) and an instruct model. When results disappoint, change one thing, train again and compare on the same held-back examples. Better data usually beats any setting.

Training loss should fall and level off. If it drops close to zero, the model is memorising: use fewer epochs or more data. If it jumps around, lower the learning rate.

Numbers help, but read the model's answers on the held-back examples, and on the tricky questions you know users will ask. Check that it still refuses what it should. See Evaluation.

A model can repeat what it was trained on. Strip API keys, passwords and personal data from training data, and keep your Langtrain API key in an environment variable, never in code or in a browser.

Smaller models are cheaper and faster to serve. Try a 3B model before an 8B, and quantise to GGUF for llama.cpp when the quality holds up. See Run your model anywhere.