| name | together-fine-tuning |
| description | LoRA, full fine-tuning, DPO preference tuning, VLM training, function-calling tuning, reasoning tuning, and BYOM uploads on Together AI. Reach for it whenever the user wants to adapt a model on custom data rather than only run inference, evaluate outputs, or host an existing model. |
Together Fine-Tuning
Overview
Use Together AI fine-tuning when the user needs to adapt a model to their own data or behavior.
Supported workflows in this repo:
- LoRA fine-tuning
- full fine-tuning
- DPO preference tuning
- VLM fine-tuning
- function-calling fine-tuning
- reasoning fine-tuning
- BYOM upload paths
When This Skill Wins
- Train a model on custom instruction or conversational data
- Improve function-calling reliability with supervised examples
- Train on preferences rather than only demonstrations
- Fine-tune multimodal or reasoning-oriented models
- Deploy a fine-tuned output model later through dedicated endpoints
Hand Off To Another Skill
- Use
together-chat-completions for plain inference without training
- Use
together-evaluations to measure a model before or after tuning
- Use
together-dedicated-endpoints to host the resulting tuned model
- Use
together-gpu-clusters only when the user needs raw infrastructure rather than managed tuning
Quick Routing
- Standard LoRA or full fine-tuning
- DPO preference tuning
- Function-calling tuning
- Reasoning tuning
- VLM tuning
- Model support and deployment options
Workflow
- Choose the tuning method that matches the desired behavior change.
- Validate dataset format before spending tokens on training.
- Upload training data and keep the returned file ID.
- Create the job with explicit method-specific parameters.
- Monitor job state, events, and checkpoints before handing off to deployment.
High-Signal Rules
- Python scripts require the Together v2 SDK (
together>=2.0.0). If the user is on an older version, they must upgrade first: uv pip install --upgrade "together>=2.0.0".
- Prefer LoRA unless the user has a specific reason to pay for full fine-tuning.
- Keep data-format validation close to the upload step so bad files fail early.
- Treat deployment as a separate phase; fine-tuning success does not automatically mean serving success.
- Use the method-specific script instead of overloading one generic workflow for all modes.
- Parameterize dataset paths, model IDs, and suffixes in automation instead of embedding one demo dataset forever.
Resource Map
Official Docs
