Base models
The encoders your model starts from, all made by our lab, and the data they were trained on.
Your model is not trained from zero: it starts from a base model, an encoder that already reads language and makes
typed decisions, and learns your decision on top of it. Every base model is ours, named TS-{base}-{size}, open and
cleared for commercial use (Apache-2.0). GET /v1/encoders lists them with what can be picked right now.
| Base model | encoder | Size | Languages | Context | Decision Index 0.3 | Training |
|---|---|---|---|---|---|---|
| TS-moBERTo-150M (recommended) | ts-moberto-150m | 150M | EN, PT | 8,192 tokens | 31.20 / 33.89 | Available |
| TS-Ettin-150M | ts-ettin-150m | 150M | EN | 8,192 tokens | 31.96 / 34.78 | Available |
| TS-Ettin-68M | ts-ettin-68m | 68M | EN | 8,192 tokens | 27.61 / 30.46 | Available |
| TS-Ettin-32M | ts-ettin-32m | 32M | EN | 8,192 tokens | 19.70 / 22.45 | Available |
| TS-Ettin-17M | ts-ettin-17m | 17M | EN | 8,192 tokens | 15.76 / 18.52 | Available |
The Decision Index 0.3 public index, in our lab's run: the encoder alone, then with the code and calculator experts that ship with the published models. Measured on public benchmarks, not on your task: your own model is measured on your examples, against Jev (the refund calculation). The context is how much text one example can hold; longer examples are refused, never cut.
TS-moBERTo-150M
Our default base, in English and Portuguese, built for typed decisions (choice, noul, score, with probabilities and
a confidence) on a CPU. It starts from moBERTo (Apache-2.0)
and was published before as Dinah-1, whose official Decision Index run
scored 34.32, with the raw results published for auditing. Held-out typed decisions (1,000 items each): 0.790 accuracy
in English, 0.780 in Portuguese.
TS-Ettin
The same recipe on the Ettin rerankers (Apache-2.0), in four sizes, English only. The 150M one is the strongest base on the index; the smaller ones trade points for speed and memory, for phones, browsers and high volumes. Their confidence is not calibrated yet; your trained model is still measured the same way.
What they were trained on
All five follow the same recipe on the same data, in two stages (2,048 tokens, then 8,192) and an average of both. Each source with the license its own card declares. Sources under CC BY-SA and CDLA-Sharing are used for training only and are not redistributed.
| Area | Sources |
|---|---|
| Typed decisions | ZefanCai/Open-Jev (CC0-1.0), LocalLLaMA/typed-decisions (Apache-2.0), typed-decisions-pt translated with Qwen3-4B (Apache-2.0), tasksource/procedural-typed-decisions (Apache-2.0) |
| Intents and customer support | AmazonScience/massive (CC BY 4.0), Bitext customer support (CDLA-Sharing-1.0), BANKING77 (CC BY 4.0), CLINC150 (CC BY 3.0), FAQ_BACEN (Apache-2.0) |
| Inference and fact checking | WANLI (CC BY 4.0), SNLI (CC BY-SA 4.0), HoVer (CC BY-SA 4.0), FEVER-NLI (MIT; FEVER and Wikipedia CC BY-SA 3.0), ContractNLI (CC BY 4.0), bAbI-NLI (BSD), FaQuAD-NLI (CC BY 4.0), ExtraGLUE RTE and COPA in Portuguese (MIT) |
| Reasoning and knowledge | CommonsenseQA (MIT), Cosmos QA (CC BY 4.0), BoolQ (CC BY-SA 3.0), HellaSwag (MIT), WinoGrande (Apache-2.0), ARC train split (CC BY-SA 4.0), QASC (CC BY 4.0), StrategyQA (MIT), BIG-bench tasks without BBH (Apache-2.0), MMLU auxiliary train, ARC and OpenBookQA parts (MIT), GSM8K (MIT), CLadder (MIT) |
| Health, finance and law | PubMedQA (MIT), MedMCQA (Apache-2.0), twitter-financial-news-sentiment (MIT), financial-tweets-sentiment (MIT), NOSIBLE financial-sentiment (ODC-By), FinQA hallucination detection (MIT), pyRis/SEntFiN (MIT), CUAD (CC BY 4.0), jurisprudencias_br (CC BY 4.0), IBM Claim Stance (CC BY 3.0) |
| Sentiment, emotion and moderation | GoEmotions (Apache-2.0), Civil Comments (CC0-1.0), iSarcasmEval (MIT), B2W-Reviews01 (CC BY 4.0), multilingual-sentiments in Portuguese (Apache-2.0) |
| Retrieval and hallucination | ESCI (Apache-2.0), HaluEval (Apache-2.0), RAGTruth (MIT), CNN/DailyMail (Apache-2.0), SQuAD 2.0 (CC BY-SA 4.0) |
| Tools | glaive-function-calling-v2 (Apache-2.0), ToolACE (Apache-2.0), hermes-function-calling-v1 (Apache-2.0), nvidia/When2Call (CC BY 4.0) |
| Phishing | LLMGen-Phishing-Email-Dataset (Apache-2.0), aura-phishing-email-corpus, SpamAssassin part (CC BY 4.0) |
| Chess | Lichess puzzles (CC0-1.0), google-deepmind/searchless_chess (Apache-2.0) |
| Humor | newyorker_caption_contest, matching task (CC BY 4.0) |
| Made by us | English to Portuguese translations with opus-mt-tc-big-en-pt (CC BY 4.0); synthetic phishing emails written with gpt-oss-120b, labels set by code (Apache-2.0); synthetic home-appliance households, RAG hallucination examples and programmatic skills (arithmetic, dates, units, sorting, graphs, code output), generated by our scripts (MIT and Apache-2.0) |
The training blocks were checked against the Decision Index's public test rows (exact and near-duplicate removal, and 8-gram overlap for the newest blocks).
Which one to pick
- Not sure, or Portuguese: TS-moBERTo-150M, the default. It is the strongest on our held-out typed decisions (0.790 in English, 0.780 in Portuguese, against 0.709 in English for TS-Ettin-150M).
- English, and you want to compare: TS-Ettin-150M, the highest on the Decision Index. Train both on the same data and keep the version that wins on your held-out examples.
- Every millisecond or megabyte counts: TS-Ettin-68M, 32M or 17M.
Every base model trains the same way and is measured by the same rule, so trying another one is a new version of the same training.