Oleh Shliazhko

Fix type checking of offload optimizer before checked class was imported

3

This bug prevents to run Megatron-LM 10B offload training example

Так работает ``` >>AddressExtractor()('проспекте ленина 39').matches [Match([MorphTagToken('проспекте', [0, 9), 'RU', 'O', [Form('проспект', Grams(NOUN,inan,loct,masc,sing))]), MorphTagToken('ленина', [10, 16), 'RU', 'I', [Form('ленин', Grams(NOUN,Sgtm,Surn,anim,gent,masc,sing)), Form('ленин', Grams(NOUN,Sgtm,Surn,accs,anim,masc,sing)), Form('ленин', Grams(NOUN,Sgtm,Surn,anim,femn,nomn,sing)), Form('ленин', Grams(ADJF,Infr,Poss,gent,masc,sing)), Form('ленин', Grams(ADJF,Infr,Poss,accs,anim,masc,sing)), Form('ленин', Grams(ADJF,Poss,femn,nomn,sing)),...

Fix setting model dropoutE when resume and remove lr overwrite

1

1. Fix setting of model.dropoute to the value from cli arguments. 2. Remove overwriting lr to initial value in resume because it leads to model divergence when training has been...

Evaluation script for Huggingface Causal models

15

It was made using the existing [FLAN eval script](https://github.com/hendrycks/test/blob/master/evaluate_flan.py) as a reference. Minor changes: - load models as Float16; - put the samples on the same device as a model;...

Speed up evaluation by caching task environments as docker images

7

#### What does this implement? This PR introduces the ability to cache the environment created for each SWE-bench task as the docker image. It saves the filesystem and environment variables...

[WIP] Browser backends

1

1. Separation of the miniwob task and underlying browser backend 2. MCP-Playwright as browser ## Description by Korbit AI ### What change is being made? Introduce a new Browser/MCP-based backend...

Configs and updates to gaia eval

1

- Use tool calling - Eval Apriel model - Added readme ## Description by Korbit AI ### What change is being made? Update Gaia evaluation configs and flow to use...