Fine-tuning and LoRA let B2B marketing teams adapt a language model to their own style, formats and classification rules. Fine-tuning trains a model further on examples; LoRA (low-rank adaptation) does this cheaply by training small adapter weights instead of the whole model. Both make sense only after prompting and retrieval have been tried, and only when you have hundreds of good examples of the output you want.
What do fine-tuning and LoRA actually change?
A base model has general ability. Fine-tuning shifts its behaviour toward your examples: tone, structure, label choices, output format. It does not reliably teach new facts. If the goal is “know our latest product specs”, retrieval-augmented generation (RAG) is the better tool, because facts change and RAG reads current documents.
LoRA is a technique from a 2021 research paper by Microsoft researchers. It freezes the original model and trains small matrices added to some layers. The adapter is a small file, can be swapped per task, and needs far less GPU memory than full fine-tuning. QLoRA goes further by working on a quantized model, which makes training possible on a single strong GPU for many open models.
When should a B2B team fine-tune?
Use a decision order. Most teams stop at step one or two.
- Prompting with clear instructions and a few examples.
- Retrieval, when the problem is missing knowledge.
- Fine-tuning, when the model knows enough but will not follow your format or style reliably, or when a smaller fine-tuned model can replace a large expensive one.
Good B2B candidates: classifying inbound leads into your own segments, tagging content by your taxonomy, writing short outputs in a strict house format such as product release notes, and extracting fields from RFPs into a fixed structure. Poor candidates: long-form thought leadership, anything needing current facts, and tasks with fewer than a few hundred examples.
Where can you fine-tune?
Hosted options include OpenAI’s fine-tuning API for selected GPT models, Google Vertex AI for Gemini and open models, Amazon Bedrock for several model families, Azure AI Foundry, and Mistral’s platform. These hide the infrastructure; you upload data and pay for training and usage.
For open-weight models such as Llama, Mistral, Qwen or Gemma, libraries like Hugging Face PEFT and TRL, Unsloth and Axolotl run LoRA training on your own or rented GPUs. This route gives more control and keeps data in your environment, at the cost of engineering time.