Prompting tells a model what to do each time. Fine-tuning changes the model so it already knows.
How it works
You start from a trained model and run more training on your own examples. In supervised fine-tuning (SFT), each example is a conversation with the answer you want. OpenAI’s guide uses JSONL, “one complete JSON structure on every line”, in the chat format:
{"messages": [{"role": "user", "content": "Classify: 'card charged twice'"}, {"role": "assistant", "content": "billing"}]}
The training job adjusts the weights so those answers become more likely, and returns a new model ID. OpenAI sets the minimum at 10 examples and recommends starting with 50 and evaluating before adding more.
Full fine-tuning updates every weight, which is costly for large models. The LoRA paper by Hu and colleagues calls deploying separate fine-tuned copies of GPT-3, “each with 175B parameters”, prohibitively expensive. Parameter-efficient methods such as LoRA train a small add-on instead. Distillation is a related idea: fine-tune a small model on a larger model’s outputs.
Why it matters
- Shorter prompts. Behaviour that needed a long few-shot prompt can move into the weights, cutting input tokens per call.
- Consistency. Good fits are classification, a fixed output style, domain translation and correcting instruction-following failures.
- Upkeep. A tuned model is pinned to one base model. When the base is retired, you retrain. OpenAI’s guide now says it is winding down its fine-tuning platform, closed to new users, with tuned models served until their base models shut down.
- Measurement first. Build an eval before training, or you cannot tell whether the tuned model is better.
Fine-tuning is a poor way to add facts that change. Retrieval keeps them current.
On BroadNet Inference
BroadNet’s API reference does not document a fine-tuning API, training jobs or uploading your own weights. It does say that “our Arabic fine-tunes and other models on request” are reached with a private model ID issued with the endpoint. BroadNet lists one such model on request: a fine-tune of Cohere Transcribe Arabic for numbers, dates and names in noisy audio. For the public Arabic transcription model, see the Arabic speech-to-text guide. If you need to train your own, an open-weight model is the usual starting point; the trade-offs are in open-weight vs closed models.