Training Jobs
Create fine-tuning jobs across 250+ models using 15+ training methods, 6 alignment algorithms, and continued pre-training. All jobs are billed per second of GPU compute.
Creating a Job
The training wizard guides you through a 7-step process:
- Model: search and select a base model from HuggingFace (e.g., Llama, Mistral, Qwen, Gemma).
- Method: choose SFT, RLHF, Continued Pre-Training (CPT), or VLM.
- Adapter: pick an adapter compatible with your model and method (LoRA, QLoRA, full fine-tune).
- Dataset: select from your uploaded datasets or import one from HuggingFace.
- GPU: choose a GPU tier based on model size and budget. The platform recommends appropriate hardware.
- Config: set hyperparameters: learning rate, epochs, batch size, LoRA rank, etc.
- Review: confirm all settings and estimated cost before launching.
Training Methods
| Method | Type | Description |
|---|---|---|
| SFT | Supervised | Standard supervised fine-tuning on instruction/response pairs |
| DPO | Alignment | Direct Preference Optimization: align the model with preference pairs |
| SimPO | Alignment | Simple Preference Optimization: reference-free alignment |
| ORPO | Alignment | Odds Ratio Preference Optimization: combined SFT + alignment |
| CPO | Alignment | Contrastive Preference Optimization with NLL regularization |
| KTO | Alignment | Kahneman-Tversky Optimization: works with binary feedback |
| CPT | Pre-training | Continued Pre-Training: extend model knowledge with domain-specific corpora |
| VLM | Multimodal | Vision-Language Model fine-tuning on image + text pairs |
i
For models under 13B parameters, LoRA is usually the best balance of quality and cost. For larger models, QLoRA lets you fine-tune on more affordable hardware.
Monitoring Jobs
Once a job is running, the training detail page shows:
- Metrics graphs: live charts for training loss, evaluation loss, learning rate schedule, and gradient norms.
- Logs: streaming stdout/stderr from the training process.
- Checkpoints: saved model snapshots at configurable intervals. Each checkpoint can be downloaded or deleted.
- Cost tracking: running cost for the current job in real time.
Stopping and Resuming
You can stop a running job at any time by clicking the Stop button on the training detail page. When you stop a job:
- Checkpoints are preserved: all saved checkpoints remain available for download.
- Billing stops immediately: you are only charged for the compute time used up to the stop point, billed per second.
- Logs and metrics are retained: the training loss graph, logs, and configuration remain accessible on the job detail page.
- Resume from checkpoint: stopped jobs can be resumed from the last saved checkpoint, continuing training where it left off.
i
If a training job looks like it has converged (loss has plateaued), you can stop it early to save costs. The checkpoints already saved are fully usable.
BiOS Documentation. Need help? Email help-bios@us.inc