Open Source

Models we built.
Open on Hugging Face.

We build models for paying customers, then we release them. 16+ models, 80K+ monthly downloads. Maintained, not abandoned.

View on Hugging Face Contribute on GitHub

Hugging Face

Search models, datasets, users...

bardsai/eu-pii-anonymization-multilang

Like42

Text ClassificationTransformersPyTorchendeesfr+17

Model card

Files and versions

Community3

Model card

Multilingual PII anonymization model for EU compliance. Detects and redacts names, emails, phones, addresses, and more across 20+ languages.

pip install transformers accelerate
from transformers import pipeline
pipe = pipeline("token-classification",
  model="bardsai/eu-pii-anonymization-multilang")

Downloads last 30 days

12,841

License: Apache 2.0

What people are actually downloading.

23+
Open models on Hugging Face
2.7M
Lifetime model downloads
8+
Languages covered

// Models

Grouped by what they do.

Each one started as work for a paying customer. We released the ones that turned out to be useful to other teams. The popular ones we keep updating.

Financial Sentiment

8 models

Language-specific models for detecting sentiment in financial news and market commentary.

finance-sentiment-ja-base
↓ 1.9M
finance-sentiment-zh-base
↓ 49K
finance-sentiment-fr-base
↓ 35K

View all 8 models

Polish Whisper

5 models

Whisper ASR models fine-tuned for Polish - winners of Hugging Face's Whisper Fine-Tuning Sprint.

whisper-large-v2-pl-v2
↓ 14K
whisper-small-pl
↓ 8.5K
whisper-medium-pl
↓ 1.9K

View all 5 models

Jaskier 7B LLM

5 models

A top-ranked ≤7B open LLM family fine-tuned with Direct Preference Optimization.

jaskier-7b-dpo
↓ 30K
jaskier-7b-dpo-v5.6
↓ 25K
jaskier-7b-dpo-v6.1
↓ 17K

View all 5 models

// Featured project

Jaxpot - JAX-native self-play RL.

Self-play RL training in JAX. We built it to train an agent for Dark Hex (a hidden-information variant of Hex). The pieces - AlphaZero-style policy/value, MCTS, vectorized environments, Hydra configs - work for any two-player adversarial game.

AlphaZero-style training loop
Vectorized JAX environments
MCTS + neural policies
Snapshot leagues & ELO
Hydra-configured experiments
Tic-tac-toe → Dark Hex

View on GitHub Read the deep-dive

Jaxpot - JAX-based self-play RL training lab for hidden games

Recent releases

Our latest open-source releases.

View all releases

Mar 16, 2026
eu-pii-anonimization-multilang
Multilingual PII detection across all official EU languages. Token-level redaction.
Token Classification
427
Feb 29, 2024
jaskier-7b-dpo-v6-GGUF
Quantized GGUF build of Jaskier 7B v6 for local inference with llama.cpp.
Text Generation
225
Feb 20, 2024
jaskier-7b-dpo-v6.1
Truthy DPO fine-tune of Jaskier 7B with improved factuality and instruction-following.
Text Generation
17K
Feb 16, 2024
jaskier-7b-dpo-v5.6
Math-focused DPO checkpoint of Jaskier 7B with stronger reasoning on numeric tasks.
Text Generation
25K

Why we release this stuff.

Survived a paying customer
Each model went through a real engagement first. The eval suite, the edge cases, and the latency target were set by the customer - not by us.
Maintained, not abandoned
The popular ones get updates. Bug reports get answered. The models we don't maintain are tagged that way so you know which are which.
Releasing pays off
80K+ monthly downloads means people who use our work end up hiring us. Open is the cheapest credibility we know how to build.

Models we built.Open on Hugging Face.