Ottema

Open, specialized AI models for Brazilian Portuguese and reliable AI systems.

We build open-source models that solve concrete production problems in Brazilian Portuguese. Our work focuses on open-vocabulary information extraction, structured-output recovery for agents, speech recognition, and operational AI for real-world workflows.

Based in Brazil. Open research, reproducible benchmarks, production-oriented models.

Brazilian Portuguese Extraction

Open-vocabulary NER and evidence extraction for real PT-BR text — including noisy operational domains where standard models fail.

Model	What it's for	Result
`ottema/gliner2-ptbr-harem` (v0.12b)	NER on journalistic/formal PT-BR. Best entity F1 among compared models on HAREM.	entity F1 = 0.4749 (macro) / 0.4501 (micro), 4x faster than BERT-CRF
`ottema/gliner2-ptbr` (v0.4)	Generalist NER for informal PT-BR (chat, atendimento, suporte).	entity F1 = 0.9976 on synthetic benchmark
`ottema/gliner2-ptbr-ontoevidence` (v0.18)	Ontology-guided evidence extraction with hard-negative rejection. First model to break the GLiNER2 "yes-man" failure mode.	F1 = 0.32 on OE test, avg 4.4 pred/text
`ottema/gliner2-ptbr-ontoevidence-data`	2268 samples, 3 splits, multi-label spans + hard negatives.	Apache-2.0

💡 Also published as ottema/gliner2-ptbr-v23 (same weights, 90+ downloads).

Browse the full Ottema Open Models collection: huggingface.co/collections/ottema/ottema/ottema-open-models-6a3600e6cc0bdc9c01dd68c8

Reliable Agents

Small specialized models that recover structured output when the LLM fails — JSON, tool calls, schema-constrained generation.

Model	What it does
`ottema/structfix-codet5p-220m`	Repairs broken JSON/tool-call output from upstream LLMs against a target schema. 220M params, fast, deterministic.
`ottema/structfix-bench`	Benchmark: 250k examples of schema-guided generation with controlled noise and constraint coverage.
`ottema/constraint-dsl`	Compact DSL for declaring typed constraints over JSON outputs.

Speech & Edge AI

Lightweight ASR and small models for resource-constrained deployment.

Model	What it's for
`ottema/stt_pt_quartznet15x5_ctc_small`	Research baseline / lightweight CPU ASR reference for PT-BR. See Nemotron full-stack for current SOTA.

How we work

Open research, reproducible benchmarks. Every model ships with training data, evaluation scripts, and ablations documented.
Production-oriented. We optimize for the metric that matters in deployment: latency, label F1 on the user's actual distribution, hard-negative robustness.
Honest about failure modes. We publish what didn't work (see the "yes-man problem" with OntoEvidence).
No private data. All training data is either synthetic, openly licensed, or used for training only (not redistributed).

Try it

ottema/gliner2-ptbr-demo — interactive Gradio demo with model selection, label presets, and 7 example sentences spanning journalistic, informal, and operational Portuguese.
ottema/structfix-demo — repair broken JSON / tool-call output against typed schemas. 5 schema presets, 10 broken-output presets, 13 paired examples.

Credits

Our models build on:

GLiNER / GLiNER2 (Urchade Zaratiana et al.) — open-vocabulary NER architecture
fastino/gliner2-multi-v1 (Fastino) — multilingual GLiNER2 base
microsoft/mdeberta-v3-base — multilingual encoder
CodeT5+ (Salesforce) — seq2seq for structured output repair
Linguateca HAREM, lfcc, arubenruben — Portuguese NER datasets (training only)

License

All models and datasets are released under Apache-2.0 unless otherwise noted.

Contact

Organization: huggingface.co/ottema