from dataclasses import dataclass from enum import Enum @dataclass(frozen=True) class Task: benchmark: str metric: str col_name: str type: str baseline: float = 0.0 # Select your tasks here # --------------------------------------------------- class Tasks(Enum): # task_key in the json file, metric_key in the json file, name to display in the leaderboard # task2 = Task("belebele_pol_Latn", "acc,none", "belebele_pol_Latn", "multiple_choice", 0.279) task3 = Task("polemo2_in", "exact_match,score-first", "polemo2-in_g", "generate_until", 0.416) task4 = Task("polemo2_in_multiple_choice", "acc,none", "polemo2-in_mc", "multiple_choice", 0.416) task5 = Task("polemo2_out", "exact_match,score-first", "polemo2-out_g", "generate_until", 0.368) task6 = Task("polemo2_out_multiple_choice", "acc,none", "polemo2-out_mc", "multiple_choice", 0.368) task7 = Task("polish_8tags_multiple_choice", "acc,none", "8tags_mc", "multiple_choice", 0.143) task8 = Task("polish_8tags_regex", "exact_match,score-first", "8tags_g", "generate_until", 0.143) task9a = Task("polish_belebele_mc", "acc,none", "belebele_mc", "multiple_choice", 0.279) task9 = Task("polish_belebele_regex", "exact_match,score-first", "belebele_g", "generate_until", 0.279) task10 = Task("polish_dyk_multiple_choice", "f1,none", "dyk_mc", "multiple_choice", 0.289) task11 = Task("polish_dyk_regex", "f1,score-first", "dyk_g", "generate_until", 0.289) task12 = Task("polish_ppc_multiple_choice", "acc,none", "ppc_mc", "multiple_choice", 0.419) task13 = Task("polish_ppc_regex", "exact_match,score-first", "ppc_g", "generate_until", 0.419) task14 = Task("polish_psc_multiple_choice", "f1,none", "psc_mc", "multiple_choice", 0.466) task15 = Task("polish_psc_regex", "f1,score-first", "psc_g", "generate_until", 0.466) task16 = Task("polish_cbd_multiple_choice", "f1,none", "cbd_mc", "multiple_choice", 0.149) task17 = Task("polish_cbd_regex", "f1,score-first", "cbd_g", "generate_until", 0.149) task18 = Task("polish_klej_ner_multiple_choice", "acc,none", "klej_ner_mc", "multiple_choice", 0.343) task19 = Task("polish_klej_ner_regex", "exact_match,score-first", "klej_ner_g", "generate_until", 0.343) task21 = Task("polish_polqa_reranking_multiple_choice", "acc,none", "polqa_reranking_mc", "multiple_choice", 0.5335588952710677) # multiple_choice task22 = Task("polish_polqa_open_book", "levenshtein,none", "polqa_open_book_g", "generate_until", 0.0) # generate_until task23 = Task("polish_polqa_closed_book", "levenshtein,none", "polqa_closed_book_g", "generate_until", 0.0) # generate_until task24 = Task("polish_poquad_open_book", "levenshtein,none", "poquad_open_book", "generate_until", 0.0) task25 = Task("polish_eq_bench_first_turn", "first_eqbench,none", "eq_bench_first_turn", "generate_until", 0.0) task26 = Task("polish_eq_bench", "average_eqbench,none", "eq_bench", "generate_until", 0.0) task20 = Task("polish_poleval2018_task3_test_10k", "word_perplexity,none", "poleval2018_task3_test_10k", "other") # task27 = Task("polish_eq_bench", "revised_eqbench,none", "eq_bench_revised", "other", 0.0) g_tasks = [task.value.benchmark for task in Tasks if task.value.type == "generate_until"] mc_tasks = [task.value.benchmark for task in Tasks if task.value.type == "multiple_choice"] rag_tasks = ['polish_polqa_reranking_multiple_choice', 'polish_polqa_open_book', 'polish_poquad_open_book'] all_tasks = g_tasks + mc_tasks NUM_FEWSHOT = 0 # Change with your few shot # --------------------------------------------------- # Your leaderboard name TITLE = """

Open PL LLM Leaderboard (0-shot and 5-shot)

Leaderboard was created as part of an open-science project SpeakLeash.org

""" # What does your leaderboard evaluate? INTRODUCTION_TEXT = f""" The leaderboard evaluates language models on a set of Polish tasks. The tasks are designed to test the models' ability to understand and generate Polish text. The leaderboard is designed to be a benchmark for the Polish language model community, and to help researchers and practitioners understand the capabilities of different models. For now, models are tested without theirs templates. Almost every task has two versions: regex and multiple choice. * _g suffix means that a model needs to generate an answer (only suitable for instructions-based models) * _mc suffix means that a model is scored against every possible class (suitable also for base models) Average columns are normalized against scores by "Baseline (majority class)". * `,chat` suffix means that a model is tested using chat templates * `,chat,multiturn` suffix means that a model is tested using chat templates and fewshot examples are treated as a multi-turn conversation * 🚧 prefix means that not all tasks were calculated and the average scores are not accurate We gratefully acknowledge Polish high-performance computing infrastructure PLGrid (HPC Centers: ACK Cyfronet AGH) for providing computer facilities and support within computational grant no. PLG/2024/016951. """ # Which evaluations are you running? how can people reproduce what you have? LLM_BENCHMARKS_TEXT = f""" ## Do you want to add your model to the leaderboard? Contact with me: [LinkedIn](https://www.linkedin.com/in/wrobelkrzysztof/) or join our [Discord SpeakLeash](https://discord.gg/FfYp4V6y3R) ## TODO * fix long model names * add inference time * add more tasks * fix scrolling on Firefox ## Tasks Tasks taken into account while calculating averages: * Average: {', '.join(all_tasks)} * Avg g: {', '.join(g_tasks)} * Avg mc: {', '.join(mc_tasks)} * Avg RAG: {', '.join(rag_tasks)} | Task | Dataset | Metric | Type | |---------------------------------|---------------------------------------|-----------|-----------------| | polemo2_in | allegro/klej-polemo2-in | accuracy | generate_until | | polemo2_in_mc | allegro/klej-polemo2-in | accuracy | multiple_choice | | polemo2_out | allegro/klej-polemo2-out | accuracy | generate_until | | polemo2_out_mc | allegro/klej-polemo2-out | accuracy | multiple_choice | | 8tags_mc | sdadas/8tags | accuracy | multiple_choice | | 8tags_g | sdadas/8tags | accuracy | generate_until | | belebele_mc | facebook/belebele | accuracy | multiple_choice | | belebele_g | facebook/belebele | accuracy | generate_until | | dyk_mc | allegro/klej-dyk | binary F1 | multiple_choice | | dyk_g | allegro/klej-dyk | binary F1 | generate_until | | ppc_mc | sdadas/ppc | accuracy | multiple_choice | | ppc_g | sdadas/ppc | accuracy | generate_until | | psc_mc | allegro/klej-psc | binary F1 | multiple_choice | | psc_g | allegro/klej-psc | binary F1 | generate_until | | cbd_mc | ptaszynski/PolishCyberbullyingDataset | macro F1 | multiple_choice | | cbd_g | ptaszynski/PolishCyberbullyingDataset | macro F1 | generate_until | | klej_ner_mc | allegro/klej-nkjp-ner | accuracy | multiple_choice | | klej_ner_g | allegro/klej-nkjp-ner | accuracy | generate_until | | polqa_reranking_mc | ipipan/polqa | accuracy | multiple_choice | | polqa_open_book_g | ipipan/polqa | levenshtein | generate_until | | polqa_closed_book_g | ipipan/polqa | levenshtein | generate_until | | poleval2018_task3_test_10k | enelpol/poleval2018_task3_test_10k | word perplexity | other | | polish_poquad_open_book | enelpol/poleval2018_task3_test_10k | levenshtein | generate_until | | polish_eq_bench_first_turn | speakleash/EQ-Bench-PL | eq_bench | generate_until | | polish_eq_bench | speakleash/EQ-Bench-PL | eq_bench | generate_until | ## Reproducibility To reproduce our results, you need to clone the repository: ``` git clone https://github.com/speakleash/lm-evaluation-harness.git -b polish3 cd lm-evaluation-harness pip install -e . ``` and run benchmark for 0-shot and 5-shot: ``` lm_eval --model hf --model_args pretrained=speakleash/Bielik-7B-Instruct-v0.1 --tasks polish_generate --num_fewshot 0 --output_path results/ --log_samples lm_eval --model hf --model_args pretrained=speakleash/Bielik-7B-Instruct-v0.1 --tasks polish_mc --num_fewshot 0 --output_path results/ --log_samples lm_eval --model hf --model_args pretrained=speakleash/Bielik-7B-Instruct-v0.1 --tasks polish_generate_few --num_fewshot 5 --output_path results/ --log_samples lm_eval --model hf --model_args pretrained=speakleash/Bielik-7B-Instruct-v0.1 --tasks polish_mc --num_fewshot 5 --output_path results/ --log_samples ``` With chat templates: ``` lm_eval --model hf --model_args pretrained=speakleash/Bielik-7B-Instruct-v0.1 --tasks polish_generate --num_fewshot 0 --output_path results/ --log_samples --apply_chat_template lm_eval --model hf --model_args pretrained=speakleash/Bielik-7B-Instruct-v0.1 --tasks polish_mc --num_fewshot 0 --output_path results/ --log_samples --apply_chat_template lm_eval --model hf --model_args pretrained=speakleash/Bielik-7B-Instruct-v0.1 --tasks polish_generate_few --num_fewshot 5 --output_path results/ --log_samples --apply_chat_template lm_eval --model hf --model_args pretrained=speakleash/Bielik-7B-Instruct-v0.1 --tasks polish_mc --num_fewshot 5 --output_path results/ --log_samples --apply_chat_template ``` ## List of Polish models * speakleash/Bielik-7B-Instruct-v0.1 * speakleash/Bielik-7B-v0.1 * Azurro/APT3-1B-Base * Azurro/APT3-1B-Instruct-v1 * Voicelab/trurl-2-7b * Voicelab/trurl-2-13b-academic * OPI-PG/Qra-1b * OPI-PG/Qra-7b * OPI-PG/Qra-13b * szymonrucinski/Curie-7B-v1 * sdadas/polish-gpt2-xl ### List of multilingual models * meta-llama/Llama-2-7b-chat-hf * mistralai/Mistral-7B-Instruct-v0.1 * HuggingFaceH4/zephyr-7b-beta * HuggingFaceH4/zephyr-7b-alpha * internlm/internlm2-chat-7b-sft * internlm/internlm2-chat-7b * mistralai/Mistral-7B-Instruct-v0.2 * teknium/OpenHermes-2.5-Mistral-7B * openchat/openchat-3.5-1210 * Nexusflow/Starling-LM-7B-beta * openchat/openchat-3.5-0106 * berkeley-nest/Starling-LM-7B-alpha * upstage/SOLAR-10.7B-Instruct-v1.0 * meta-llama/Llama-2-7b-hf * internlm/internlm2-base-7b * mistralai/Mistral-7B-v0.1 * internlm/internlm2-7b * alpindale/Mistral-7B-v0.2-hf * internlm/internlm2-1_8b """ EVALUATION_QUEUE_TEXT = """ ## Some good practices before submitting a model ### 1) Make sure you can load your model and tokenizer using AutoClasses: ```python from transformers import AutoConfig, AutoModel, AutoTokenizer config = AutoConfig.from_pretrained("your model name", revision=revision) model = AutoModel.from_pretrained("your model name", revision=revision) tokenizer = AutoTokenizer.from_pretrained("your model name", revision=revision) ``` If this step fails, follow the error messages to debug your model before submitting it. It's likely your model has been improperly uploaded. Note: make sure your model is public! Note: if your model needs `use_remote_code=True`, we do not support this option yet but we are working on adding it, stay posted! ### 2) Convert your model weights to [safetensors](https://huggingface.co/docs/safetensors/index) It's a new format for storing weights which is safer and faster to load and use. It will also allow us to add the number of parameters of your model to the `Extended Viewer`! ### 3) Make sure your model has an open license! This is a leaderboard for Open LLMs, and we'd love for as many people as possible to know they can use your model 🤗 ### 4) Fill up your model card When we add extra information about models to the leaderboard, it will be automatically taken from the model card ## In case of model failure If your model is displayed in the `FAILED` category, its execution stopped. Make sure you have followed the above steps first. If everything is done, check you can launch the EleutherAIHarness on your model locally, using the above command without modifications (you can add `--limit` to limit the number of examples per task). """ CITATION_BUTTON_LABEL = "Copy the following snippet to cite these results" CITATION_BUTTON_TEXT = r""" @misc{open-pl-llm-leaderboard, title = {Open PL LLM Leaderboard}, author = {Wróbel, Krzysztof and {SpeakLeash Team} and {Cyfronet Team}}, year = 2024, publisher = {Hugging Face}, howpublished = "\url{https://huggingface.co/spaces/speakleash/open_pl_llm_leaderboard}" } """