How to fine-tune Olmo 3 (LoRA, QLoRA)

This guide covers adapting the Olmo 3 family with LoRA or QLoRA. Estimate cost with the training cost calculator and see fine-tuning for the overview.

Which checkpoint

• Base (Olmo-3-1025-7B) was trained on Dolma 3 in three stages: pretraining, mid-training and long-context training. Its card says you can fine-tune from the final checkpoint or from intermediate revisions named stage1-stepXXX, stage2-stepXXX and stage3-stepXXX, using OLMo-core scripts or other recipes.
• Think and Instruct models were post-trained with SFT, DPO and RLVR. The cards point to the Open-Instruct repository for the DPO and RLVR stages.
• For a light domain adapter, start from Instruct (or Think if you need reasoning). For your own full post-training recipe, start from Base.

Chat template

Use the tokenizer's template with <|im_start|> and <|im_end|>. Keep the system message style of the checkpoint you start from. Think data should keep reasoning inside <think> tags.

Target modules

The cards do not list LoRA target modules. The 7B base has 32 layers, hidden size 4096 and 32 query and key-value heads, so inspect the module names in the loaded model.

License

The code and weights are Apache 2.0, which allows derivative models. The cards add that the models are intended for research and educational use in accordance with Ai2's Responsible Use Guidelines, so read those before shipping a commercial derivative.