Skip to content

Latest commit

 

History

History
105 lines (85 loc) · 6.41 KB

readme.md

File metadata and controls

105 lines (85 loc) · 6.41 KB
Raw

Evaluation scripts for MMLU

Claude

The commands below run Claude models on the MMLU dataset.

The engine 'claude-v1.3' is the full capacity one, whereas 'claude-instant-v1.0' is the lightweight version.

You can also specify prompt type to either 'single' or 'multiple'. 'single' refers to a group of demonstration questions with answers being supplied as a part of the prompt, whereas 'multiple' refers to having rounds of conversations for demonstration in the prompt.

cd MMLU
mkdir outputs
API_KEY=<your_api_key>
python run_mmlu_claude.py --anthropic_key=${API_KEY} --engine=claude-v1.3 --prompt_type='multiple'

LLaMA v.s. Falcon

The commands below run open-resource models on the MMLU dataset (answer-only setting, no CoT).

We have evaluated the LLaMA-65B and Falcon-40B. The data and the prompts come from the MMLU original repo. We use greedy decoding. For LLaMA we use fp16. For Falcon we evaluate both fp16 and bf16 (and the performance is similar).

LLAMA_CKPT_DIR=<path to model checkpoints>
PARAM_SIZE=65 # 7, 13, 33, 65
MODEL_TYPE=llama # ["llama", "falcon"] 
python run_mmlu_open_source.py --ckpt_dir ${LLAMA_CKPT_DIR} --param_size ${PARAM_SIZE} --model_type ${MODEL_TYPE}

Previously there was a bug incurred by long prompts, resulting LLaMA getting 0 scores on high_school_european_history and high_school_us_history. Under that commit, LLaMA average score is 61.44.

Now we have updated the code by popping out in-context examples to make the prompt fit into the context length (for us and eu history). After fixing this, the LLaMA score becomes 63.64, basically the same number as the one reported in the paper.

Then using the identical script, Falcon 40B gets 49.08.

Below is the detailed evaluation results of the LLaMA family and Falcon family on each subtask.

Task LLaMA-65B Falcon-40B (FP16) Falcon-40B (BF16) LLaMA-33B LLaMA-13B LLaMA-7B Falcon-7B (BF16) LLaMA-65B (Bug)
Average 0.6364 0.4996 0.4996 0.5847 0.4696 0.3522 0.2641 0.6144
abstract_algebra 0.3000 0.2900 0.2800 0.3700 0.3000 0.2600 0.2700 0.3000
anatomy 0.5778 0.4963 0.4963 0.5185 0.4667 0.3852 0.3185 0.5778
astronomy 0.7303 0.4803 0.4868 0.6118 0.4671 0.3421 0.3026 0.7303
business_ethics 0.5900 0.5800 0.5800 0.5500 0.4600 0.4100 0.2100 0.5900
clinical_knowledge 0.6604 0.5472 0.5509 0.5736 0.4604 0.3396 0.2566 0.6604
college_biology 0.6944 0.5556 0.5486 0.5903 0.4722 0.3750 0.2014 0.6944
college_chemistry 0.4800 0.2900 0.3000 0.4500 0.3000 0.3000 0.2300 0.4800
college_computer_science 0.4700 0.4600 0.4700 0.4600 0.4100 0.3000 0.2600 0.4700
college_mathematics 0.3500 0.3200 0.3300 0.3600 0.3700 0.3500 0.3000 0.3500
college_medicine 0.5376 0.4162 0.4220 0.5491 0.4162 0.3353 0.2717 0.5376
college_physics 0.3529 0.2059 0.2157 0.2549 0.1961 0.2353 0.2059 0.3529
computer_security 0.8000 0.5500 0.5600 0.6700 0.6400 0.4500 0.3200 0.8000
conceptual_physics 0.5830 0.4213 0.4170 0.5064 0.3830 0.3745 0.2596 0.5830
econometrics 0.3947 0.2807 0.2719 0.3596 0.2895 0.2807 0.2368 0.3947
electrical_engineering 0.5517 0.5172 0.5172 0.5310 0.4483 0.2207 0.3310 0.5517
elementary_mathematics 0.3968 0.3095 0.3069 0.3677 0.2698 0.2646 0.2751 0.3968
formal_logic 0.4365 0.2619 0.2540 0.3571 0.3175 0.2619 0.1508 0.4365
global_facts 0.3800 0.3300 0.3200 0.3800 0.3600 0.3000 0.3500 0.3800
high_school_biology 0.7452 0.6065 0.5935 0.6903 0.5129 0.3323 0.2387 0.7452
high_school_chemistry 0.4089 0.3547 0.3448 0.4138 0.2808 0.2759 0.2956 0.4089
high_school_computer_science 0.6900 0.5000 0.4900 0.6100 0.5000 0.3400 0.3000 0.6900
high_school_european_history 0.7939 0.5394 0.5455 0.7515 0.6182 0.4182 0.3091 0.0000
high_school_geography 0.7879 0.6768 0.6616 0.7121 0.5505 0.3384 0.2576 0.7879
high_school_government_and_politics 0.8808 0.7098 0.7047 0.8290 0.6684 0.4560 0.1762 0.8808
high_school_macroeconomics 0.6564 0.5026 0.5026 0.5667 0.4590 0.3436 0.2077 0.6590
high_school_mathematics 0.3444 0.2630 0.2667 0.2778 0.2593 0.2519 0.2407 0.3407
high_school_microeconomics 0.6807 0.4706 0.4706 0.5798 0.4748 0.3151 0.2563 0.6765
high_school_physics 0.3642 0.3179 0.3113 0.3510 0.3113 0.2781 0.2715 0.3709
high_school_psychology 0.8257 0.6514 0.6422 0.7596 0.6128 0.4826 0.2275 0.8257
high_school_statistics 0.6157 0.3981 0.4028 0.4907 0.2963 0.3333 0.1991 0.6204
high_school_us_history 0.8333 0.6225 0.6225 0.7892 0.5882 0.3676 0.2745 0.0000
high_school_world_history 0.8354 0.5316 0.5274 0.8017 0.6751 0.4346 0.2911 0.8354
human_aging 0.6726 0.6368 0.6323 0.6771 0.5336 0.3946 0.3453 0.6726
human_sexuality 0.7786 0.6336 0.6107 0.6489 0.5725 0.3511 0.2824 0.7786
international_law 0.8182 0.6694 0.6612 0.7521 0.6446 0.5124 0.3058 0.8182
jurisprudence 0.7407 0.6296 0.6296 0.7130 0.5093 0.4167 0.2778 0.7500
logical_fallacies 0.7730 0.5828 0.5828 0.6810 0.5276 0.4172 0.2515 0.7730
machine_learning 0.4732 0.2679 0.2857 0.4107 0.3214 0.2768 0.3482 0.4732
management 0.8252 0.7476 0.7476 0.7670 0.6602 0.3398 0.2330 0.8252
marketing 0.8718 0.7821 0.7778 0.8419 0.7179 0.4615 0.2821 0.8718
medical_genetics 0.6800 0.6300 0.6100 0.6600 0.5300 0.3700 0.2700 0.6800
miscellaneous 0.8135 0.6909 0.6884 0.7778 0.6475 0.4253 0.2835 0.8135
moral_disputes 0.7370 0.5809 0.5925 0.6676 0.5000 0.4075 0.2746 0.7370
moral_scenarios 0.4782 0.2872 0.2994 0.4045 0.2961 0.2425 0.2637 0.4771
nutrition 0.6895 0.5850 0.5817 0.6242 0.5196 0.3889 0.2549 0.6895
philosophy 0.7363 0.6013 0.6045 0.6624 0.5531 0.4051 0.2894 0.7363
prehistory 0.7407 0.5833 0.5895 0.6821 0.5185 0.3457 0.2716 0.7377
professional_accounting 0.4858 0.4326 0.4326 0.4468 0.3475 0.2660 0.2660 0.4858
professional_law 0.4980 0.3768 0.3781 0.4648 0.3709 0.3018 0.2373 0.4935
professional_medicine 0.6213 0.4596 0.4632 0.5588 0.5074 0.4412 0.1838 0.6176
professional_psychology 0.6667 0.4788 0.4788 0.6373 0.4869 0.3529 0.2794 0.6650
public_relations 0.7545 0.6455 0.6545 0.6909 0.6091 0.4000 0.2545 0.7545
security_studies 0.7184 0.5755 0.5673 0.6653 0.5306 0.3347 0.2490 0.7224
sociology 0.8109 0.7114 0.7114 0.7910 0.6119 0.4577 0.2886 0.8109
us_foreign_policy 0.8800 0.7700 0.7700 0.8300 0.7800 0.4300 0.3900 0.8800
virology 0.5301 0.4036 0.4217 0.4940 0.4337 0.3434 0.3916 0.5301
world_religions 0.8129 0.7953 0.7953 0.7953 0.6784 0.4912 0.3743 0.8129