Serbian LLM eval 🇷🇸

Note: it can likely also be used for other HBS languages (Croatian, Bosnian, Montenegrin) - support for these languages is on my roadmap (see future work).

What is currently covered:

• Common sense reasoning: Hellaswag, Winogrande, PIQA, OpenbookQA, ARC-Easy, ARC-Challenge
• World knowledge: NaturalQuestions, TriviaQA
• Reading comprehension: BoolQ

You can find the Serbian LLM eval dataset on HuggingFace. For more details on how the dataset was built see this technical report on Weights & Biases. The branch serbevaltranslate was used to do machine translation, while serbevalrefine was used to do further refinement using GPT-4.

Please email me at gordicaleksa at gmail com in case you're willing to sponsor the projects I'm working on.

You will get the credits and eternal glory. :)

In Serbian: ``` I na srpskom, ukoliko ste voljni da finansijski podržite ovaj poduhvat korišćenja ChatGPT da se dobiju kvalitetniji podaci, i koji je od nacionalnog/regionalnog interesa, moj email je gordicaleksa at gmail com. Dobićete priznanje na ovom projektu da ste sponzor (i postaćete deo istorije). :)

Dalje ovaj projekat će pomoći da se pokrene lokalni large language model ekoksistem. ``````

Run the evals

Step 1. Create Python environment

git clone https://github.com/gordicaleksa/lm-evaluation-harness-serbian cd lm-evaluation-harness-serbian pip install -e .

Currently you might need to manually install also the following packages (do pip install): sentencepiece, protobuf, and one more (submit PR if you hit this).

Step 2. Tweak the launch json and run

--model_args <- any name from HuggingFace or a path to HuggingFace compatible checkpoint will work

--tasks <- pick any subset of these arc_challenge,arc_easy,boolq,hellaswag,openbookqa,piqa,winogrande,nq_open,triviaqa

--num_fewshot <- set the number of shots, should be 0 for all tasks except for nq_open and triviaqa (these should be run in 5-shot manner if you want to compare against Mistral 7B)

--batch_size <- depending on your available VRAM set this as high as possible to get the max speed up

Future work:

• Cover popular aggregated results benchmarks: MMLU, BBH, AGI Eval and math: GSM8K, MATH
• Explicit support for other HBS languages.

Sponsors

Thanks to all of our sponsor(s) for donating for the yugoGPT (first 7B HBS LLM) & Serbian LLM eval projects.

yugoGPT base model will soon be open-source under permissive Apache 2.0 license.

Platinum sponsors

• Ivan (anon)

Gold sponsors

• qq (anon)
• Adam Sofronijevic
• Yanado
• Mitar Perovic
• Nikola Ivancevic
• Rational Development DOO
• Ivan i Natalija Kokić

Silver sponsors

• psk.rs
• OmniStreak
• Luka Važić
• Miloš Durković
• Marjan Radeski
• Marjan Stankovic
• Nikola Stojiljkovic
• Mihailo Tomić
• Bojan Jevtic
• Jelena Jovanović
• Nenad Davidović
• Mika Tasich
• TRENCH-NS
• Nemanja Grujičić
• Mladen Fernežir
• tim011

Also a big thank you to the following individuals: - Slobodan Marković - for spreading the word! :) - Aleksander Segedi - for help around bookkeeping

Credits

A huge thank you to the following technical contributors who helped translate the evals from English into Serbian: * Vera Prohaska * Chu Kin Chan * Joe Makepeace * Toby Farmer * Malvi Bid * Raphael Vienne * Nenad Aksentijevic * Isaac Nicolas * Brian Pulfer * Aldin Cimpo

License

Apache 2.0

Citation

@article{serbian-llm-eval, author = "Gordić Aleksa", title = "Serbian LLM Eval", year = "2023" howpublished = {\url{https://huggingface.co/datasets/gordicaleksa/serbian-llm-eval-v1}}, }