We train your model. You own the result.
Our ML engineers handle data preparation, training, alignment, evaluation and deployment on GREENAIBOX clusters. You receive weights, code, recipes and a benchmark report, with full IP transfer and no lock-in.
Fine-tuning sprints from $4,900. Foundation model pre-training on up to 4,096 GPUs.
From a weekend fine-tune to a frontier pre-train
Fine-tuning
SFT, LoRA and QLoRA, or full-parameter fine-tunes of open-weight models on your data. Instruction, chat, code, domain adaptation and multimodal.
Alignment
DPO, RLHF, GRPO and constitutional methods with reward models built from your preference data. Safety evals included.
Continued pre-training
Extend a base model with billions of domain tokens: legal, medical, financial, scientific or your own codebase and documentation.
Pre-training from scratch
Tokenizer, data mixture, curriculum and architecture design, then distributed training on 256 to 4,096 GPUs with fault-tolerant checkpointing.
Distillation and quantization
Compress large models into small ones that fit your latency and cost targets: distillation, pruning, FP8, INT4 and speculative decoding.
Evaluation and deployment
Held-out evals, standard benchmarks, red-teaming and an optimized inference container. Hosted on GREENAIBOX or handed over to your platform.
How a sprint runs, week by week
You see everything as it happens: the experiment tracker, the daily standup and the eval dashboard are yours from day one.
- 01
Scoping
Define the task, the evaluation set that defines success, the base model and the compute budget. You get a written plan with milestones and a fixed price.
- 02
Data
Audit for quality, duplication, leakage and licensing. Clean, format and document. You approve the dataset before any GPU spins up.
- 03
Training
Hyperparameter sweeps on a slice, then the full run on a reserved cluster. Checkpoints, metrics and logs stream to a workspace you can open any time.
- 04
Evaluation and handover
Results on your eval set, standard benchmarks and the base model. Weights, code, recipe, data pipeline and inference container delivered. Optional hosting.
Fixed scope, fixed price
Compute is billed transparently at published rates. Engineering is priced per engagement so you know the total before you start.
Fine-tuning sprint
Adapt an open-weight model to your data in about four weeks.
- Data audit, cleaning and formatting
- SFT, LoRA / QLoRA or full fine-tune of models up to 70B
- Held-out evaluation set and benchmark report
- Weights, training code and recipe delivered to you
Custom training program
A dedicated ML team and reserved cluster for continuous model development.
- Everything in the sprint tier
- Alignment: DPO, RLHF, constitutional and reward modeling
- Distillation, quantization and inference optimization
- Weekly evals, experiment tracking and reporting
- Named ML lead and infrastructure engineer
Foundation model pre-training
Pre-train from scratch on 256 to 4,096 GPUs with full ownership of the result.
- Tokenizer, data mixture and curriculum design
- Distributed training on InfiniBand clusters
- Checkpoint management and fault-tolerant runs
- Full IP transfer: weights, data pipeline and code
Built for teams that can't afford a leak
- Single-tenant clusters in the region you choose, or on your own hardware
- Named engineers only. Access is scoped per project and fully audited
- Encryption at rest, in transit and for every checkpoint
- Your DPA, your terms. HIPAA, GDPR and ITAR-constrained engagements supported
- Full IP transfer of weights, code, data pipelines and evals at handover
- Deletion on request with a signed certificate of destruction
What lands in your repo at handover
Model weights
Final and best checkpoints in safetensors, plus quantized variants.
Training code and recipe
Reproducible configs, scripts and environment so you can rerun or extend the work.
Data pipeline and data card
Cleaning, filtering and formatting code with documentation of every decision.
Evaluation report
Results on your eval set, public benchmarks and safety tests, with error analysis.
Inference container
vLLM or TensorRT-LLM image tuned for your latency and throughput targets.
Runbook
How to serve, monitor, retrain and roll back, written for your team.
Knowledge transfer
Recorded walkthrough sessions and two weeks of post-handover support.
IP assignment
Signed transfer of all artifacts and a certificate of data deletion on request.
Who owns the model and the code?
You do, fully. Every engagement includes IP transfer of weights, training code, data pipelines, evaluation harnesses and recipes. We retain nothing except what you explicitly allow us to reference in aggregate.
How do you protect our data?
Training runs on single-tenant clusters in the region you choose, with encryption at rest and in transit, access restricted to the named engineers on your project, and full audit logs. We sign your DPA and can work under HIPAA, GDPR or ITAR constraints. On-premises engagements on your own cluster are also available.
Do we need to have an ML team on our side?
No. You need someone who understands the business problem and can judge whether the outputs are good. We handle everything technical and explain decisions in plain language.
What does compute cost on top of the engagement fee?
Sprints bill compute at published GPU Cloud rates. Programs bill a reserved cluster separately at published GPU Cluster rates. Pre-training is quoted as a single price covering compute and engineering, based on token budget and cluster size.
Can you deploy and host the model afterwards?
Yes. We hand over an optimized inference container (vLLM or TensorRT-LLM) that runs anywhere. If you want us to host it, we serve it on GREENAIBOX with autoscaling and an SLA.
How do you measure success?
Before training starts we agree on an evaluation set and target metrics. We report against that set, against standard benchmarks and against the base model. If the sprint misses the agreed target, we run another iteration at no engineering charge.
Tell us what you want the model to do
Send a paragraph about your use case and data. We'll come back within a day with a recommended approach, timeline and fixed price.