PruhaNLP

Small open models and the benchmarks we needed to check them. Mostly Russian NLP, lately robots and microcontrollers.

Robotics

• RoboSim at Home — browser studio for the SO-100 arm: collect with a leader arm, train, RL, eval in MuJoCo.
• TurboVLA-base — vision-language-action model without an LLM. 0.2B params, ~40 ms per action chunk, ~1 GB VRAM.
• Demo policies on one pick-and-place task: TurboVLA · SmolVLA · ACT · dataset

On-device

• needle-3-esp32-s3 — tool-calling model running offline on an ESP32-S3. Text in, schema-valid JSON out.

1C code search

• USER2-1C-code — first open embedding model for 1C:Enterprise / BSL code.
• 1C-Ebench — code retrieval benchmark built from real developer questions.
• 1C-Code-Train — 784K query / code / hard-negative triplets.

Russian NLP

• ModernMT-en-ru-EXP — 66M-param EN→RU translator, on par with NLLB-1.3B on FLORES-200. Demo.
• RuSlangX — do LLMs understand Russian internet slang? Scores range from 24% to 97%.