What is USF BIOS: The Complete AI Training Platform

USF BIOS is a fully managed AI training platform purpose-built for teams that need to fine-tune large language models without managing the complexity of distributed training infrastructure. BIOS handles cluster setup, dependency management, checkpoint storage, and job orchestration for you, so machine learning engineers can spend their time curating datasets, designing training runs, and evaluating model quality.

The platform supports over 250 pre-trained models across more than 80 architectures, spanning the major open-weight model families released in the past several years. Instruction-tuned chat models, base models for continued pre-training, and multimodal vision-language models all train through the same interface. This breadth of support means you can experiment across model families without switching platforms or rebuilding tooling.

BIOS is designed for a broad range of use cases: adapting a general-purpose model to a specialized vertical, teaching a model a proprietary output format, injecting domain knowledge from internal documentation, or aligning model behavior to match company values and safety standards. All of these workloads fit naturally into the BIOS training pipeline.

The defining feature of BIOS is its depth of training method support. Most fine-tuning platforms offer a small selection of adapter-based methods, typically LoRA or QLoRA alongside basic supervised fine-tuning. BIOS goes substantially further, offering 15 or more distinct training methods covering the full spectrum from lightweight parameter-efficient adapters all the way to full parameter fine-tuning and continued pre-training.

On the alignment side, BIOS supports six preference optimization algorithms (DPO, SimPO, ORPO, CPO, KTO, and reward modeling), allowing teams to move beyond simple instruction-following and actually shape model behavior to match human preferences. This matters in production, where output quality, safety, and adherence to organizational standards decide whether a model ships.

Dataset management gets the same attention as training itself. The platform validates datasets at upload time, provides previews so you can inspect samples before training begins, and supports direct import from HuggingFace Hub for both public and private datasets. Real-time training metrics are streamed to the dashboard throughout every run, giving you visibility into loss curves, evaluation performance, and checkpoint health without needing to SSH into a machine or tail a log file.

Here is the BIOS dashboard showing the central control panel where you can monitor all your training activity, wallet balance, and recent jobs at a glance:

BIOS covers every major open-weight model family in active use today. In the Llama family, you can train on Llama 3, Llama 3.1, Llama 3.2, and Llama 3.3 in their various parameter sizes, giving you access to Meta's latest architectures with their improved instruction-following and extended context capabilities. The Qwen family from Alibaba Cloud is fully supported, including Qwen 2 and Qwen 2.5 in both base and instruction-tuned variants, as well as the Qwen-VL vision-language models.

Mistral AI's models, including the Mistral base models and Mixtral mixture-of-experts variants, are available for fine-tuning on BIOS. The DeepSeek family, including DeepSeek V2 and V3 with their efficient multi-head latent attention architecture, is supported as well. Microsoft's Phi family, including Phi 3, Phi 3.5, and Phi 4, gives teams access to high-quality small language models suitable for edge or low-latency deployments. Google's Gemma 2 rounds out the dense transformer coverage.

For vision-language modeling, BIOS supports InternVL 2 and InternVL 2.5, the LLaVA family of visual instruction models, MiniCPM-V for compact multimodal workloads, and DeepSeek-VL. This multimodal coverage is rare among fine-tuning platforms and positions BIOS as a strong choice for teams building applications that need to understand both text and images. The breadth of supported architectures, over 80 in total, means that as new model families emerge, BIOS is positioned to add support rapidly.

BIOS organizes its 15+ training methods into several categories. At the supervised fine-tuning layer, you can run standard SFT for instruction-response datasets, or opt for continued pre-training when you want to extend the model's knowledge base with unlabeled text corpora before applying instruction tuning on top.

Parameter-efficient adapter methods include LoRA (Low-Rank Adaptation), QLoRA (quantized LoRA for memory efficiency), AdaLoRA (which adaptively allocates the rank budget across layers), LoHa (which uses Hadamard products for a different form of low-rank parameterization), BOFT (Butterfly Orthogonal Fine-Tuning for structured weight updates), and ReFT (Representation Fine-Tuning, which operates on hidden representations rather than weight matrices). Having all these adapters in one place lets you empirically compare them on your specific task and dataset rather than relying on general recommendations.

On the alignment side, the six preference optimization methods each take a different approach to encoding human preference signals into the model. DPO uses a binary cross-entropy loss directly on chosen and rejected response pairs. SimPO removes the need for a reference model and uses sequence-average log probabilities. ORPO combines supervised fine-tuning and preference learning in a single pass. CPO uses a contrastive objective to keep chosen response probabilities high. KTO (Kahneman-Tversky Optimization) uses a prospect-theory-inspired loss that works even without paired preference data. Reward modeling trains a separate scalar scoring model that can then be used in downstream reinforcement learning pipelines. Online reinforcement learning methods including PPO, GRPO, and GKD are also on the roadmap and coming soon.

One of the most practically significant aspects of BIOS is its per-second billing model. Nearly every alternative platform in the fine-tuning space bills by the hour, meaning that a job lasting 35 minutes is billed as a full hour. For teams running many short experiments, which is exactly what good machine learning practice requires, hourly billing inflates costs substantially and discourages the kind of rapid iteration that leads to better models.

With per-second billing, you pay only for the actual wall-clock compute time your training job consumes. A 22-minute job costs you exactly 22 minutes. A hyperparameter sweep with 12 short trial runs costs you the sum of their actual durations, not 12 rounded-up hours. This billing granularity is particularly valuable in the early stages of a project when you are running quick experiments to validate your dataset, tune your learning rate, or compare adapter types before committing to a longer full training run.

These savings show up in practice. Teams that migrate from hourly platforms to BIOS frequently report meaningful reductions in monthly training spend, not because they are running fewer experiments, but because they are no longer paying for compute they did not use. Budgeting gets easier too. You can estimate costs from wall-clock time rather than having to round up every job to the next hour boundary.

Getting started with BIOS follows a straightforward four-step workflow designed to minimize friction between your raw dataset and a trained model. First, sign up for an account and create your workspace. BIOS handles workspace provisioning automatically, including storage allocation for datasets and checkpoints. There is no infrastructure to configure and no environment to set up. Every new account receives a $10 welcome credit so you can start training immediately without entering payment information.

Second, upload your dataset. BIOS accepts JSONL files with standard instruction-tuning or preference-pair schemas, and validates the format at upload time so you catch errors before they abort a training run. If your data lives on HuggingFace Hub, you can import it directly by connecting your HuggingFace account and selecting the dataset from the BIOS interface. The dataset preview feature lets you inspect samples to confirm the formatting looks correct.

Third, configure your training run. Select the base model you want to fine-tune from the 250+ available models, choose your training method (SFT, LoRA, QLoRA, DPO, or any of the other 15+ options), and set your hyperparameters. BIOS provides sensible defaults for all hyperparameters and includes inline documentation to help you understand the tradeoffs. You can also name your run and add tags for organizational purposes.

Fourth, launch the job and monitor it from the dashboard. Training metrics stream in real time, including training loss, validation loss, learning rate, and throughput. When the job completes, your model checkpoints are saved and available for download or for serving through compatible inference infrastructure. The entire cycle from dataset to trained model can be completed in minutes for small datasets, or hours for large-scale full fine-tuning runs.

One of the most underappreciated aspects of a good training platform is how it handles datasets. Poor dataset tooling leads to wasted compute: jobs that fail 20 minutes in because of a formatting error, or that produce poor models because the data contained unexpected noise that was not visible during preparation.

BIOS treats dataset management as a core workflow. When you upload a dataset, the platform immediately validates the file format, checks for required fields, and reports any schema violations before your data is stored. This validation catches the most common mistakes: missing instruction fields, misnamed columns, invalid JSON lines, and encoding issues. Catching these errors at upload time rather than at training time saves both compute and frustration.

The dataset preview feature goes a step further. After upload, you can browse individual samples from your dataset directly in the browser. This lets you visually confirm that the data looks as expected: that instructions are well-formed, that responses are the right length, and that any metadata fields are populated correctly. For preference datasets, you can inspect both the chosen and rejected examples side by side.

BIOS also supports importing datasets directly from HuggingFace Hub. If you have connected your HuggingFace account through the Integrations page, you can browse both public and your private datasets, preview them, and import with a single click. This eliminates the download-convert-upload cycle that is common on other platforms and lets you experiment with publicly available datasets before investing in your own data curation.

BIOS is built for teams, not just individual practitioners. The platform uses a two-level hierarchy of organizations and workspaces that maps naturally to how most ML teams are structured. A personal organization is created automatically when you sign up, and team access is managed through invitations. Admins can invite members by email, assign roles (Owner, Admin, or Member), and create separate workspaces for different projects or teams.

Workspaces provide logical isolation for datasets, training jobs, and model checkpoints. A research team running exploratory experiments can work in one workspace while a production team running deployment-ready training runs works in another, all within the same organization. Each workspace has its own member list, so you can control who has access to sensitive datasets or production models.

Billing is managed at the organization level through a shared wallet. This centralized billing model simplifies cost tracking. You can see total spend across all workspaces and team members in a single view, making it easy to allocate costs to projects or departments.