oz-eval / README.md
Stopwolf's picture
Add Hermes3, Llama3.1 70B
3c02b90 verified
|
raw
history blame
5.64 kB
metadata
language:
  - sr
license: apache-2.0
task_categories:
  - question-answering
dataset_info:
  features:
    - name: index
      dtype: int64
    - name: questions
      dtype: string
    - name: options
      sequence: string
    - name: answer
      dtype: string
    - name: answer_index
      dtype: int64
  splits:
    - name: test
      num_bytes: 200846
      num_examples: 1003
  download_size: 139630
  dataset_size: 200846
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/oz-eval-*

OZ Eval

image/png

Dataset Description

OZ Eval (sr. Opšte Znanje Evaluacija) dataset was created for the purposes of evaluating General Knowledge of LLM models in Serbian language. Data consists of 1k+ high-quality questions and answers which were used as part of entry exams at the Faculty of Philosophy and Faculty of Organizational Sciences, University of Belgrade. The exams test the General Knowledge of students and were used in the enrollment periods from 2003 to 2024.

This is a joint work with @Stopwolf!

Evaluation process

Models are evaluated by the following principle using HuggingFace's library lighteval. We supply the model with the following template:

Pitanje: {question}

Ponuđeni odgovori:
A. {option_a}
B. {option_b}
C. {option_c}
D. {option_d}
E. {option_e}

Krajnji odgovor:

We then compare likelihoods of each letter (A, B, C, D, E) and calculate the final accuracy. All evaluations are ran in a 0-shot manner using a chat template.

GPT-like models were evaluated by taking top 20 probabilities of the first output token, which were further filtered for letters A to E. Letter with the highest probability was taken as a final answer.

Exact code for the task can be found here.

You should run the evaluation with the following command (do not forget to add --use_chat_template):

accelerate launch lighteval/run_evals_accelerate.py \
    --model_args "pretrained={MODEL_NAME},trust_remote_code=True" \
    --use_chat_template \
    --tasks "community|serbian_evals:oz_task|0|0" \
    --custom_tasks "/content/lighteval/community_tasks/oz_evals.py" \
    --output_dir "./evals" \
    --override_batch_size 32

Evaluation results

Model Size Accuracy Stderr
GPT-4-0125-preview ??? 0.9199 ± 0.002
GPT-4o-2024-05-13 ??? 0.9196 ± 0.0017
GPT-3.5-turbo-0125 ??? 0.8245 ± 0.0016
Llama3.1-70B-Instruct [4bit] 70B 0.8185 ± 0.0122
GPT-4o-mini-2024-07-18 ??? 0.7971 ± 0.0005
Mustra-7B-Instruct-v0.2 7B 0.7388 ± 0.0098
Tito-7B-slerp 7B 0.7099 ± 0.0101
Yugo55A-GPT 7B 0.6889 ± 0.0103
Zamfir-7B-slerp 7B 0.6849 ± 0.0104
Mistral-Nemo-Instruct-2407 12.2B 0.6839 ± 0.0104
Llama-3.1-SauerkrautLM-8b-Instruct 8B 0.679 ± 0.0147
Qwen2-7B-instruct 7B 0.6730 ± 0.0105
Llama-3-SauerkrautLM-8b-Instruct 8B 0.661 ± 0.0106
Yugo60-GPT 7B 0.6411 ± 0.0107
DeepSeek-V2-Lite-Chat 15.7B 0.6047 ± 0.0109
Llama3.1-8B-Instruct 8B 0.5972 ± 0.0155
Llama3-70B-Instruct [4bit] 70B 0.5942 ± 0.011
Hermes-3-Theta-Llama-3-8B 8B 0.5932 ± 0.0155
Hermes-2-Theta-Llama-3-8B 8B 0.5852 ± 0.011
Mistral-7B-Instruct-v0.3 7B 0.5753 ± 0.011
openchat-3.6-8b-20240522 8B 0.5513 ± 0.0111
Llama3-8B-Instruct 8B 0.5274 ± 0.0111
Starling-7B-beta 7B 0.5244 ± 0.0112
Hermes-2-Pro-Mistral-7B 7B 0.5145 ± 0.0112
Qwen2-1.5B-Instruct 1.5B 0.4506 ± 0.0111
Perucac-7B-slerp 7B 0.4247 ± 0.011
Phi-3-mini-128k-instruct 3.8B 0.3719 ± 0.0108
SambaLingo-Serbian-Chat 7B 0.2802 ± 0.01
Gemma-2-9B-it 9B 0.2193 ± 0.0092
Gemma-2-2B-it 2.6B 0.1715 ± 0.0084

Citation

@article{oz-eval,
  author    = "Stanivuk Siniša & Đorđević Milena",
  title     = "OZ Eval: Measuring General Knowledge Skill at University Level of LLMs in Serbian Language",
  year      = "2024"
  howpublished = {\url{https://huggingface.co/datasets/DjMel/oz-eval}},
}