Answer.AI and Johns Hopkins University
Train ModernBERT and Ettin decision models on your data
Encoders that score the candidate answers of typed decisions, trained and served as decision models. Supported by the training and serving software Tensorant runs. A run trained kev and a deployment served it on 24 October 2026, on an earlier version of that software, so start with a trial.
- Status
- Supported
- Input
- Text
- Kind
- Dense
- Trains with
- LoRA, full
- Suggested sizes
- LoRA from1 GPU, about 7 GB
Sizes
Sizes and GPUs
Each model with the GPUs one machine needs to train and to serve it. The figures are the estimates Tensorant checks a model against; a trial measures the real use. See GPU memory.
| Model | Parameters | Train with LoRA | Train with QLoRA | Serve | Status |
|---|---|---|---|---|---|
| answerdotai/ModernBERT-baseNew head | 150M | 1 GPUabout 7 GB | - | - | Supported |
| jhu-clsp/ettin-encoder-400mNew head | 400M | 1 GPUabout 8 GB | - | - | Supported |
| skundu42/kevDecision scorer, head kept | 396M | 1 GPUabout 8 GB | - | 1 GPUabout 4 GB in all | Supported |
- Training needs the memory shown on each GPU: more GPUs on a machine train faster, but each holds the whole model.
- Serve: the numbers of GPUs of one machine the base model can be split across, and the memory it needs in all.
Training
How it trains
Train ModernBERT and Ettin encoders with LoRA or full training, to score the candidate answers of typed decisions.
- A decision model trains on one GPU, with LoRA or in full.
- kev keeps its one-score head and its temperature, and training continues from them. The Ettin encoder and ModernBERT base get a new one-score head.
- A run scores the model it starts from and the one it trains on the version's test split, so it needs no baseline.
Every training run, trials included, pays a training-token fee for the tokens it trains on, by the model's size. Decision models are on every organization.
Serving
How it serves
Publish a trained decision model under a name, and your application asks it on the API's decisions route, POST /v1/decisions, for a probability for each candidate. See Serve decision models.
skundu42/kevalso serves as it is.jhu-clsp/ettin-encoder-400mandanswerdotai/ModernBERT-baseserve once trained, with their new head.- A published decision model shares a GPU with your other decision models of the same kind, GPU and availability.
Steps
Train ModernBERT and Ettin on Tensorant
Create a decision project
Create a decision project with
skundu42/kevas its base model, or pick another encoder from the list's decision models.Import your cases
Import typed decisions or labelled texts as sources.
Review and freeze a version
Review the cases and approve the ones to train on, then freeze them into a version with training, validation and test splits.
Run a trial
A trial trains a small sample of your cases and tells you what the full run will need. LoRA on
skundu42/kevneeds 1 GPU, about 8 GB.Train and score
A decision run scores the base model on the test split, trains, fits the temperature, and scores the tuned model on the same split.
Publish and call it
Publish the trained model under a name, and ask it with
POST /v1/decisions.
Fine-tuning
Train a calibrated decision model with kev
Fine-tune kev, a 400M ModernBERT scorer, on typed decisions, check calibration quality, and gate the run with measured scores and cost.
Deployment
Serve a decision model and call it from your app
Deploy a trained decision model, call POST /v1/decisions with batches and retries, and act on calibrated probabilities with measured cost.
Docs
Supported models
Every family's architecture, status and notes, and what a base model needs.
Start with one dataset.
Create an account and import your data, or book a demo and we will show you the product.