1. Overview
2. Quick Start
3. Leaderboards and how to interpret them
   • Models
   • Evaluators
4. Use-cases
   • Evaluating a model
   • Making a new leaderboard
   • Making a new evaluator
5. Contributing
   • Contributing a model
   • Contributing an evaluator
   • Contributing an eval set
   • Contributing a completion function
6. Limitations
7. Analysis
   • Length-controlled AlpacaEval
   • Analyzing an evaluator
   • Analyzing an eval set
8. Citation
9. Additional information
   • Length-controlled win rates
   • AlpacaEval 2.0
   • Data Release
   • Differences with AlpacaFarm
   • Related work
   • Interpreting annotations
   • Major updates

When to use AlpacaEval? Our automatic evaluator is a quick and cheap proxy for human evaluation of simple instruction-following tasks. It is useful if you have to run many evaluations quickly, e.g., during model development.

When not to use AlpacaEval? As any other automatic evaluator, AlpacaEval should not replace human evaluation in high-stake decision-making, e.g., to decide on model release. In particular, AlpacaEval is limited by the fact that (1) the instructions in the eval set might not be representative of advanced usage of LLMs; (2) automatic evaluators may have biases such as favoring style over factuality of the answer; and (3) AlpacaEval does not measure the risks that a model could cause. Details in limitations.

SYNOPSIS
    alpaca_eval COMMAND

COMMANDS
    COMMAND is one of the following:

     evaluate
       Evaluate a model based on its outputs. This is the default entrypoint if no command is specified.

     evaluate_from_model
       Evaluate a model from HuggingFace or an API provider. This is a wrapper around `evaluate` which includes generating from a desired model.

     make_leaderboard
       Precompute and save an entire leaderboard for a given dataset / evaluator / set of models generations.

     analyze_evaluators
       Analyze an evaluator and populates the evaluators leaderboard (agreement with human, speed, price,...).

Win Rate: the win rate measures the fraction of time the model's output is preferred over the reference's outputs (test-davinci-003 for AlpacaEval and gpt4_turbo for AlpacaEval 2.0). More specifically, to compute the win rate we collect pairs of outputs of the desired model on every instruction from the ApacaEval dataset. We then pair each output with the output of our reference model (e.g. text-davinci-003) on the same instruction. We then ask our automatic evaluator which output they prefer. See AlpacaEval's and AlpacaEval 2.0's prompts and configs, in particular we randomize the order of outputs to avoid position bias. We then average the preferences over all instructions in the dataset to get the win rate of the model over the baseline. If both outputs are exactly the same we use a half preference for both models.

Standard error: this is the standard error (normalized by N-1) of the win rate, i.e., the preferences averaged over the different instructions.

Our alpaca_eval_gpt4 ( see configs) annotator averages over preferences, where preferences are obtained as follows:

1. it takes in an instruction and a pair of outputs (from the desired model and the reference model)
2. if a preference was this triple was already computed, it returns it (i.e. it uses caching)
3. it randomizes the order of the outputs to avoid position bias
4. it formats the instruction and outputs into the following zero-shot prompt, which asks to order the outputs in order of preference
5. it completes the prompt using GPT4 with temperature=0
6. it parses the preference from the completions and returns it

The annotator is a mix between (and was highly influenced by) AlpacaFarm and Aviary evaluators. In particular, we use the same code as for AlpacaFarm (caching/randomization/hyperparameters) but use a ranking prompt similar to that of Aviary. We make changes to Aviary's prompt to decrease the bias for longer outputs. Details in Related work.

For AlpacaEval 2.0 we use weighted_alpaca_eval_gpt4_turbo, which uses logprobs to compute continuous preference and uses GPT4_turbo as model ( see configs).

We now explain in words how we compute the metrics in the table above. The code is here.

Human agreement: this measures the agreement between the current annotator and the majority preferences of humans on our ~650 annotations from our cross-annotation set, which contains 4 human annotations per example. To estimate the agreement between a single human (humans row in the table above) and the majority of humans, we take one of the 4 annotations and compute the accuracy that it has when predicting the mode of the other 3 annotations. We then average this accuracy over all 4 annotations and over the 650 instructions to get the human agreement, i.e., we compute the expected (over humans and samples) leave-one-out agreement. If the mode is not unique, we take one of the modes at random. We perform exactly the same computation for the automatic annotators, so that the final numbers are comparable.

Price [$/1000 examples]: this is the average price of every 1000 annotations. For humans, it is the price that we paid Mechanical Turkers to collect those annotations ($21/hour). If the price depends on the machine used to compute the annotations (e.g. Guanaco) we leave it empty.

Time [seconds/1000 examples]: this is the average time it takes to compute 1000 annotations. For humans, it is the estimated median time that each Mechanical Turker took to annotate 1000 examples. For automatic annotators, it is the average time that it took us when running the annotations. Note that this can depend on API limits that are different for different users and the number of requests that the clusters are processing.

Spearman corr.: this measures the Spearman correlation between a leaderboard computed with the auto-annotator's preference and the leaderboard computed with human preferences. As with Human agreement, we use the human annotations from AlpacaFarm but we now consider the method-level agreement rather than only the sample-wise agreement with humans. Note that we only use have 9 models and so the correlation is not very reliable.

Pearson corr.: same as with Spearman corr. but with Pearson correlation.

Bias: agreement between the most likely human label and the most likely automatic one. For automatic annotators we estimate it by sampling 4 different annotations for each example. The randomness here comes from the order of the outputs in the prompt, sampling from the LLM, and if applicable the order of the instruction in the batch and the choice of annotator in the pool. We then take the mode of the 4 annotations and compute the accuracy of the mode when predicting the mode of the 4 human annotations. Note that this is likely an overestimate on the real bias that we would get if we had an "infinite" number of cross-annotations. A low bias means that the annotator has in expectation the same preferences as humans. For the case of humans, the bias is zero by definition. Note that this is related to but not the standard statistical bias, because we take the mode instead of average over annotations and we consider 0-1 loss instead of squared loss.

Variance: expected agreement a single automatic preference and the most likely one. We estimate it the same way as we estimated "human agreement" for humans, i.e., we take the expected leave one out error when predicting the mode of the 3 annotations using the 4th annotation. A low variance means that the annotator is consistent with its preference, i.e., if you sample from it with different seeds it will give the same result. As with the bias, this is not exactly the standard statistical variance, because we take the mode instead of average over annotations and we consider 0-1 loss instead of squared loss.

Note that the "human agreement" is tightly related to the bias and variance. In particular, the variance measures the error due to the fact that we only use a single annotation while the bias aims to measure the irreducible error for the current annotator.

Proba. prefer longer: this is the probability that the annotator prefers the longer output when one of the two outputs is significantly longer than the other (more than 30 characters difference).

In the full table we also provide the following metrics:

Proba. prefer lists: this is the probability that the annotator prefers the output that contains a list/bullet points when one output does but not the other.

Proba. prefer 1: this is the probability that the annotator prefers the first of the pair of outputs. All our proposed annotators randomize over outputs in the prompt, so this should be 0.5. Prior annotators, such as lmsys and aviary, do not.

# parsed: this is the number of examples that the annotator was able to parse.

Note that if the variance and bias is empty, it means that we only performed one single annotation for each 648 example due to resource (time and price) constraints. This explains why the #parsed is 648, otherwise it should be 2592.

Overall we recommend using annotators_config=weighted_alpaca_eval_gpt4_turbo if you want the high agreement with humans, and annotators_config=chatgpt_fn if you are on a tight budget.

When choosing an annotator we recommend you to consider the following (the first three are obvious):

• "Human agreement [%]"
• "Price [$/1000 examples]"
• "Time [seconds/1000 examples]"
• "* corr." approx. > 0.7. It is important that the correlation is not too low, but we do not recommend using it as the main metric as the correlation is computed on only 9 models.
• "Proba. prefer longer" approx. < 0.7. Indeed, we found see that the majority of preference of human annotators have strong bias for longer answers (as shown by the high performance=62.2 of the "longest" evaluator that always prefers the longest output). This suggests that it might more of a bias with the human annotators. In order to avoid having leaderboards with strong biases for length, we suggest using automatic annotators with less than 0.7 "Proba. prefer longer".
• "Variance" approx. < 0.2. We believe that a good evaluator should have as little variance as possible so that results are mostly reproducible. Note that variance can be desirable in the case where we are simulating humans as shown in AlpacaFarm.

We filtered the annotators that do not satisfy those requirements in the table above (besides humans / ChatGPT / 003 / lmsys for reference purposes). For all results see here. In general, we found weighted_alpaca_eval_gpt4_turbo to be a good trade-off between quality / price / time / variance / length bias.

NAME
    alpaca_eval evaluate - Evaluate a model based on its outputs. This is the default entrypoint if no command is specified.

SYNOPSIS
    alpaca_eval evaluate <flags>

DESCRIPTION
    Evaluate a model based on its outputs. This is the default entrypoint if no command is specified.

FLAGS
    --model_outputs=MODEL_OUTPUTS
        Type: Optional[Union]
        Default: None
        The outputs of the model to add to the leaderboard. Accepts data (list of dictionary, pd.dataframe, datasets.Dataset) or a path to read those (json, csv, tsv) or a function to generate those. Each dictionary (or row of dataframe) should contain the keys that are formatted in the prompts. E.g. by default `instruction` and `output` with optional `input`. If None, we just print the leaderboard.
    -r, --reference_outputs=REFERENCE_OUTPUTS
        Type: Union
        Default: <func...
        The outputs of the reference model. Same format as `model_outputs`. If None, the reference outputs are a specific set of Davinci 003 outputs on the AlpacaEval set:
    --annotators_config=ANNOTATORS_CONFIG
        Type: Union
        Default: 'alpaca_eval_gpt4_turbo_fn'
        The path the (or list of dict of) the annotator's config file. For details see the docstring of `PairwiseAnnotator`.
    -n, --name=NAME
        Type: Optional[Optional]
        Default: None
        The name of the model to add to the leaderboard. If None we check if `generator is in model_outputs` if not we use "Current model".
    -o, --output_path=OUTPUT_PATH
        Type: Union
        Default: 'auto'
        Path to the directory where the new leaderboard and the annotations should be stored. If None we don't save. If `auto` we use `model_outputs` if it is a path, and otherwise use the directory from which we call the script.
    -p, --precomputed_leaderboard=PRECOMPUTED_LEADERBOARD
        Type: Union
        Default: 'auto'
        The precomputed leaderboard or a path to it (json, csv, or tsv). The leaderboard should contain at least the column `win_rate`. If `auto` we will try to use the corresponding leaderboard for the reference outputs (only if in CORRESPONDING_OUTPUTS_LEADERBOARDS). If `None` we won't add other models from the leaderboard.
    --is_overwrite_leaderboard=IS_OVERWRITE_LEADERBOARD
        Type: bool
        Default: False
        Whether to overwrite the leaderboard if the model is already in it.
    -l, --leaderboard_mode_to_print=LEADERBOARD_MODE_TO_PRINT
        Type: Optional
        Default: 'minimal'
        The mode of the leaderboard to use. Only used if the precomputed leaderboard has a column `mode`, in which case it will filter the leaderboard by this mode. If None keeps all.
    -c, --current_leaderboard_mode=CURRENT_LEADERBOARD_MODE
        Type: str
        Default: 'community'
        The mode of the leaderboard for the current method.
    --is_return_instead_of_print=IS_RETURN_INSTEAD_OF_PRINT
        Type: bool
        Default: False
        Whether to return the metrics instead of printing the results.
    -f, --fn_metric=FN_METRIC
        Type: Union
        Default: 'pairwise_to_winrate'
        The function or function name in `metrics.py` that will be used to convert preference to metrics. The function should take a sequence of preferences (0 for draw, 1 for base win, 2 when the model to compare wins) and return a dictionary of metrics and the key by which to sort the leaderboard.
    -s, --sort_by=SORT_BY
        Type: str
        Default: 'win_rate'
        The key by which to sort the leaderboard.
    --is_cache_leaderboard=IS_CACHE_LEADERBOARD
        Type: Optional[Optional]
        Default: None
        Whether to save the result leaderboard to `precomputed_leaderboard`. If None we save only if max_instances not None. A preferred way of adding models to the leaderboard is to set `precomputed_leaderboard` to the previously saved leaderboard at `<output_path>/leaderboard.csv`.
    --max_instances=MAX_INSTANCES
        Type: Optional[Optional]
        Default: None
        The maximum number of instances to annotate. Useful for testing.
    --annotation_kwargs=ANNOTATION_KWARGS
        Type: Optional[Optional]
        Default: None
        Additional arguments to pass to `PairwiseAnnotator.annotate_head2head`.
    -A, --Annotator=ANNOTATOR
        Default: <class 'alpaca_eval.annotators.pairwise_evaluator.PairwiseAn...
        The annotator class to use.
    Additional flags are accepted.
        Additional arguments to pass to `PairwiseAnnotator`.

NAME
    alpaca_eval evaluate_from_model - Evaluate a model from HuggingFace or an API provider. This is a wrapper around `evaluate` which includes generating from a desired model.

SYNOPSIS
    alpaca_eval evaluate_from_model MODEL_CONFIGS <flags>

DESCRIPTION
    Evaluate a model from HuggingFace or an API provider. This is a wrapper around `evaluate` which includes generating from a desired model.

POSITIONAL ARGUMENTS
    MODEL_CONFIGS
        Type: Union
        A dictionary or path (relative to `models_configs`) to a yaml file containing the configuration of the model to decode from. If a directory,we search for 'configs.yaml' in it. The keys in the first dictionary should be the generator's name, and the value should be a dictionary of the generator's configuration which should have the

FLAGS
    -r, --reference_model_configs=REFERENCE_MODEL_CONFIGS
        Type: Optional[Union]
        Default: None
        Same as in `model_configs` but for the reference model. If None, we use the default Davinci003 outputs.
    -e, --evaluation_dataset=EVALUATION_DATASET
        Type: Union
        Default: <func...
        Path to the evaluation dataset or a function that returns a dataframe. If None, we use the default evaluation
    -a, --annotators_config=ANNOTATORS_CONFIG
        Type: Union
        Default: 'alpaca_eval_gpt4_turbo_fn'
        Path to the annotators configuration or a dictionary. If None, we use the default annotators configuration.
    -o, --output_path=OUTPUT_PATH
        Type: Union
        Default: 'auto'
        Path to save the generations, annotations and leaderboard. If auto saves at `results/<model_name>`
    -m, --max_instances=MAX_INSTANCES
        Type: Optional[int]
        Default: None
        Maximum number of instances to generate and evaluate. If None, we evaluate all instances.
    --is_strip_output=IS_STRIP_OUTPUT
        Type: bool
        Default: True
        Whether to strip trailing and leading whitespaces from the outputs.
    --is_load_outputs=IS_LOAD_OUTPUTS
        Type: bool
        Default: True
        Whether to try to load outputs from the output path. If True and outputs exist we only generate outputs for instructions that don't have outputs yet.
    -c, --chunksize=CHUNKSIZE
        Type: int
        Default: 64
        Number of instances to generate before saving. If None, we save after all generations.
    Additional flags are accepted.
        Other kwargs to `evaluate`

NOTES
    You can also use flags syntax for POSITIONAL ARGUMENTS

• Caching: by default all annotations are cached on disk at caching_path. Annotations are thus never recomputed, which makes annotations faster, cheaper and allow for reproducibility. This helps even when evaluating different models as many models have the same outputs.
• Output randomization by default, we randomize over the examples of outputs, as we found that annotators tend to prefer the first examples they see.
• Batching we provide code and examples to batch annotations, which decreases cost and time for annotations if the prompt is long. See for example alpaca_farm_greedy_gpt4.
• Pool of annotators we provide code and examples to evaluate using a pool of automatic annotators, which is helpful for replicating the variance of human annotations. See for example alpaca_farm.
• Seeding based on instructions For reproducibility and more fair comparison between models, we seed all randomness (output order, order in batches, examples for each annotator in a pool) based on the instruction.

NAME
    alpaca_eval make_leaderboard - Precompute and save an entire leaderboard for a given dataset / evaluator / set of models generations.

SYNOPSIS
    alpaca_eval make_leaderboard <flags>

DESCRIPTION
    Precompute and save an entire leaderboard for a given dataset / evaluator / set of models generations.

FLAGS
    --leaderboard_path=LEADERBOARD_PATH
        Type: Optional[Union]
        Default: None
        The path to save the leaderboard to. The leaderboard will be saved as a csv file, if it already exists it will
    --annotators_config=ANNOTATORS_CONFIG
        Type: Union
        Default: 'alpaca_eval_gpt4_turbo_fn'
        The path the (or list of dict of) the annotator's config file.
    --all_model_outputs=ALL_MODEL_OUTPUTS
        Type: Union
        Default: <fu...
        The outputs of all models to add to the leaderboard. Accepts data (list of dictionary, pd.dataframe, datasets.Dataset) or a path to read those (json, csv, tsv potentially with globbing) or a function to generate those. If the path contains a globbing pattern, we will read all files matching the pattern and concatenate them. Each dictionary (or row of dataframe) should contain the keys that are formatted in the prompts. E.g. by default `instruction` and `output` with optional `input`. It should also contain a column `generator` with the name of the current model.
    -r, --reference_outputs=REFERENCE_OUTPUTS
        Type: Union
        Default: <func...
        The outputs of the reference model. Same format as `all_model_outputs` but without needing `generator`. By default, the reference outputs are the 003 outputs on AlpacaEval set.
    -f, --fn_add_to_leaderboard=FN_ADD_TO_LEADERBOARD
        Type: Callable
        Default: 'evaluate'
        The function to use to add a model to the leaderboard. If a string, it should be the name of a function in `main.py`. The function should take the arguments: `model_outputs`, `annotators_config`, `name`, `precomputed_leaderboard`, `is_return_instead_of_print`, `reference_outputs`.
    --leaderboard_mode=LEADERBOARD_MODE
        Type: str
        Default: 'verified'
        The mode of the leaderboard to save all new entries with.
    -i, --is_return_instead_of_print=IS_RETURN_INSTEAD_OF_PRINT
        Type: bool
        Default: False
        Whether to return the metrics instead of printing the results.
    Additional flags are accepted.
        Additional arguments to pass to `fn_add_to_leaderboard`.

If you want to make a new leaderboard using a single command (rather than multiple alpaca_eval calls), for your desired evaluation set and evaluators, you can use the following:

alpaca_eval make_leaderboard \
  --leaderboard_path <path_to_save_leaderboard> \
  --all_model_outputs <model_outputs_path> \
  --reference_outputs <reference_outputs_path> \
  --annotators_config <path_to_config.yaml>

where:

• leaderboard_path: path to save the leaderboard to. The leaderboard will be saved as a csv file, if it already exists it will append.
• all_model_outputs : The json path to the outputs of all models to add to the leaderboard (as a single file or by globbing multiple files). Each dictionary should contain the keys (instruction and output) that are formatted in the prompts and a column generator with the name of the current model. As an example see this file.
• reference_outputs the path to the outputs of the reference model. Each dictionary should contain the keys (instruction and output) that are formatted in the prompts. By default, the reference outputs are the 003 outputs on AlpacaEval set.
• annotators_config: The path to the annotator's config file. Defaults to alpaca_eval_gpt4.

NAME
    alpaca_eval analyze_evaluators - Analyze an evaluator and populates the evaluators leaderboard (agreement with human, speed, price,...).

SYNOPSIS
    alpaca_eval analyze_evaluators <flags>

DESCRIPTION
    Analyze an evaluator and populates the evaluators leaderboard (agreement with human, speed, price,...).

FLAGS
    --annotators_config=ANNOTATORS_CONFIG
        Type: Union
        Default: 'alpaca_eval_gpt4_turbo_fn'
        The path the (or list of dict of) the annotator's config file.
    -A, --Annotator=ANNOTATOR
        Default: <class 'alpaca_eval.annotators.pairwise_evaluator.PairwiseAn...
        The annotator class to use.
    --analyzer_kwargs=ANALYZER_KWARGS
        Type: Optional[Optional]
        Default: None
        Additional arguments to pass to the analyzer.
    -p, --precomputed_leaderboard=PRECOMPUTED_LEADERBOARD
        Type: Union
        Default: PosixPath('/Users/yanndubois/Desktop/GitHub/alpaca_eval/src/...
        The precomputed (meta)leaderboard of annotators or a path to it (json, csv, or tsv).
    --is_save_leaderboard=IS_SAVE_LEADERBOARD
        Type: bool
        Default: False
        Whether to save the leaderboard (ie analyzed results).
    --is_return_instead_of_print=IS_RETURN_INSTEAD_OF_PRINT
        Type: bool
        Default: False
        Whether to return the leaderboard (ie analyzed results). If True, it will not print the results.
    --is_overwrite_leaderboard=IS_OVERWRITE_LEADERBOARD
        Type: bool
        Default: False
        Whether to overwrite the leaderboard if it already exists.
    -m, --max_instances=MAX_INSTANCES
        Type: Optional[Optional]
        Default: None
        The maximum number of instances to analyze.
    --is_single_annotator=IS_SINGLE_ANNOTATOR
        Type: bool
        Default: False
        Whether to analyze a single annotator. If True, will not be able to estimate the annotator's bias.
    -l, --leaderboard_mode_to_print=LEADERBOARD_MODE_TO_PRINT
        Type: str
        Default: 'minimal'
        The mode of the leaderboard to print.
    -c, --current_leaderboard_mode=CURRENT_LEADERBOARD_MODE
        Type: str
        Default: 'minimal'
        The mode of the leaderboard to save all new entries with.
    -o, --output_path=OUTPUT_PATH
        Type: Union
        Default: 'auto'
        Path to save the leaderboard and annotataions. If None, we don't save.
    Additional flags are accepted.
        Additional arguments to pass to `Annotator`.

AlpacaEval provides a simple way of making new evaluators. All you need is to make a new configs.yaml configuration file, which you will then pass as --annotators_config <path_to_config.yaml> to alpaca_eval. Here are some ways you can make a new evaluator:

• Changing the prompt: Write a new prompt in a text file and specify the path in prompt_template of the configuration file. Paths are relative to the configuration file.
• Changing decoding parameters: Specify the desired parameters in completions_kwargs in the configuration file. To see all available parameters refer to the docstrings of the corresponding function in this file specified by fn_completions in the configuration file.
• Changing the model: Specify the desired model in model_name and the corresponding prompt in prompt_template. If the model comes from another provider you will have to change fn_completions which maps to the corresponding function in this file. We provide fn_completions functions to use models from OpenAI, Anthropic, Cohere, or HuggingFace. To install packages needed for all providers use pip install alpaca_eval[all].

Parameters
----------
prompt_template : path
    A prompt that will be given to `fn_prompter` or path to the prompts. Path is relative to
    `evaluators_configs/`

fn_completion_parser : callable or str
    Function in `completion_parsers.py` to use for parsing the completions into preferences. For each completion,
    the number of preferences should be equal to the batch_size if not we set all the preferences in that batch to
    NaN.

completion_parser_kwargs : dict
    Kwargs for fn_completion_parser.

fn_completions : callable or str
    Function in `decoders.py` to use for decoding the output.

completions_kwargs : dict
    kwargs for fn_completions. E.g. model_name, max_tokens, temperature, top_p, top_k, stop_seq.

is_randomize_output_order : bool
    Whether to randomize output_1, output_2 when formatting.

batch_size : int
    Number of examples that will be added in a single prompt.

Once you made the evaluator you can also analyze it and add it to the evaluator's leaderboard using the following command:

alpaca_eval analyze_evaluators --annotators_config '<path_to_config.yaml>'    

To estimate the bias and variance this evaluates every example with 4 seeds, i.e., 2.5K evaluation. If you want a cheaper evaluation you can use a single seed using --is_single_annotator True which will skip the estimation of bias and variance.

A verified result in AlpacaEval indicates that a core maintainer has decoded the outputs from the model and performed the evaluation. Unfortunately, we, the AlpacaEval maintainers, lack the resources to verify all the models and so we will only do that for models that are in the top-5 of the leaderboard. We apologize for any inconvenience this may cause and appreciate your understanding. To have your model verified, please follow the steps below:

1. Contact @yann on Discord, or email us if you have our email, providing a brief rationale for why your model should be verified.
2. Await our response and approval before proceeding.
3. Prepare a script to decode from your model that does not require a GPU, typically the same script used for your model contribution. It should run using alpaca_eval evaluate_from_model --model_configs '<your_model_name>' without requiring a local GPU.
4. Generate temporary API keys for running the script and share them with us. Specifically, we need the keys for both decoding your model and for evaluation (e.g., OpenAI or Anthropic key).
5. We will execute alpaca_eval evaluate_from_model --model_configs '<your_model_name>', update the results, and inform you so that you can revoke the temporary keys.

Note that we will not re-evaluate the same model. Due to sampling variance, the results might slightly differ from your initial ones. We will replace your previous community results with the verified ones.

Please first follow the directions in Making a new evaluator. Once you're created the annotator config, we ask that you create a new leaderboard for the annotator by evaluating the minimal set of models. The outputs for these models can be found by downloading alpaca_eval_all_outputs.json.

alpaca_eval make_leaderboard \
  --leaderboard_path src/alpaca_eval/leaderboards/data_AlpacaEval/<evaluator>_leaderboard.csv \
  --all_model_outputs alpaca_eval_all_outputs.json \
  --annotators_config <evaluator_config>

Then, please create a PR with the annotator config and leaderboard csv.

To contribute a new eval set, you'll first need to specify a set of textual instructions. Then, you'll need to specify a set of reference outputs (model win-rates are computed against this reference). For ease of use, you may use the default text-davinci-003 reference config.

Place these together into a json, where each entry specifies the fields instruction, output, and generator. You can look to alpaca_eval.json as a guide (the dataset field is not necessary).

Finally, we ask that you create a minimal leaderboard on this new evaluation set. You can do this with the following:

alpaca_eval make_leaderboard \
  --leaderboard_path <src/alpaca_eval/leaderboards/data_AlpacaEval/your_leaderboard_name.csv> \
  --all_model_outputs alpaca_eval_all_outputs.json \
  --reference_outputs <path_to_json_file>

Please submit a PR with the eval set json and corresponding leaderboard csv.

Currently, we allow different completion functions, e.g., openai, anthropic, huggingface_local, huggingface_hub_api ... If you want to contribute a new completion function / API with which to perform inference then follow those steps:

1. add a file .py with a function <name>_completions(prompts : Sequence[str], model_name :str, ... ) in the decoder folder. This function should take as argument the prompts + kwargs and return the completions. Please look at other completion functions in the directory for templates. E.g. huggingface_local_completions or anthropic.
2. add <name>_completions and dependencies in init . Again you can follow the example of huggingface_local_completions
3. update optional dependencies in setup.py
4. add a model you want to evaluate in the models configs
5. evaluate your model using alpaca_eval evaluate_from_model --model_configs '<model_configs>'
6. (optional) push the results from the previous model on AlpacaEval leaderboard following those steps

Feel free to start a PR early, we'll be able to provide some help in the process!

First, our validation of evaluators based on human cross-annotations suffers from the following limitations: (1) we qualitatively found that our crowd-workers tend to also favor style such as length and presence of lists over factuality; (2) this does not validate whether win-rates against a reference model is a good evaluation strategy in the first place; (3) preferences from 16 crowd-workers are not representative of preferences of all humans.

Second, our suggested approach to selecting evaluation sets based on statistical power suffers from the following limitations: (1) statistical power does not ensure the right direction, e.g. you can have an unnatural set of instructions where Alpaca "performs" better than better model; and (2) this can push users to select data to support the hypothesis that they want to validate.

Caution: all the following results are about AlpacaEval 1.0 and have not been updated since

Analyzing evaluators:

As we saw in the evaluator's leaderboard, there are many metrics to consider when selecting an evaluator, e.g. the quality, price, and speed. To assist with selection of the evaluator we provide a few functions to plot those metrics. The following shows for example the price/time/agreement of the different evaluators.

Here we see that alpaca_eval_gpt4 performs very well and is better than humans on all the considered metrics.

Previously we only considered the agreement with human annotators overall. An additional validation that one could do is checking whether making a leaderboard using our automatic annotator gives similar results as a leaderboard from humans. To enable such analysis, we release human annotations of outputs from 22 methods from AlpacaFarm => 22*805 = ~18K annotations. As a result we can test the correlation between the win-rates of the 22 models as evaluated by the humans and our automatic annotator. Note that this is arguably a better way of selecting an automatic evaluator than using "human agreement [%]" but is expensive given that it requires 18K annotations. The plot below shows such correlation for the alpaca_eval_gpt4 evaluator.

We see that the alpaca_eval_gpt4 leaderboard is highly correlated (0.94 Pearson correlation) to the leaderboard from humans, which further suggests that automatic evaluation is a good proxy for human evaluation. For the code and more analysis, see this notebook, or the colab notebook above.

Caution: all the following results are about AlpacaEval 1.0 and have not been updated since.

Making evaluation sets:

When creating an evaluation set there are two main factors to consider: how much data to use? and what data?

One way of answering those question is by considering a leaderboard of models that you believe are of different quality and checking what and how much data is needed to distinguish between them in a statistically significant way. We will do so below using a paired t-test to test if the difference in win-rates between every pair of models is statistically significant.

First, let us consider the question of how much data to use. Below we show the number of random samples needed from AlpacaEval for the paired t-test to give a p-value < 0.05 for each pair of models in the minimal alpaca_eval_gpt4 leaderboard. Grey cells correspond to pairs that are not significantly different on the 805 samples. y- and x-axis are ordered by the win-rate of the first and second model respectively.

We see that most models can already be distinguished with 50 samples, and that 150 samples allows distinguishing the majority of pairs (74 out of 78). This suggests that we can decrease the evaluation set size by a factor of 4 when testing two models that have similar performance gaps as those on the minimal alpaca_eval_gpt4 leaderboard.

The second question is what data to use. Again we can try to answer this question from a statistical power perspective: what data allows to best distinguish between models. Let's consider this for all the datasets that are part of AlpacaEval, but let us control for the size of the evaluation sets as we only care about the quality of the data. The following plot shows the p-values from the paired t-test of each pairs of models on 80 examples of each subset of AlpacaEval.

We see for example that the self-instruct dataset yields the least statistical power, which suggests that one could remove this dataset from the evaluation set. The exact reason should be analyzed in future work. For the code and more analysis see this notebook, or the colab notebook above.

Length controlled (LC) win-rates are a debiased version of the win-rates that control for the length of the outputs.

The main idea is that for each model we will fit a logistic regression to predict the preference of the autoannotator given: (1) the instruction, (2) the model, and (3) the difference of length between the baseline and model output. Given such a logistic regression we can then try to predict the counterfactual "what would the preference be if the model's output had the same length as the baseline" by setting the length difference to 0. By averaging over this length-controlled preference, we then obtain the length-controlled win-rate. The exact form of the logistic regression is taken such that the interpretation of LC win rates is similar to the raw win rates, for example for any model m1 and m2 we have win_rate(m1, m2) = 1 - win_rate(m2, m1) \in [0,100] and win_rate(m1, m1) = 0.5. Length controlled win-rates increase the correlation between AlpacaEval's leaderboard and Chat Arena from 0.93 to 0.98 Spearman correlation, while significantly decreasing the length gameability of the annotator. For more information and results about length controlled win-rates see this notebook.

This idea of estimating the controlled direct effect, by predicting the outcome while conditioning on the mediator (the length difference), is common in statistical inference.

To get LC win rates on previously annotated models, you can use the following command:

pip install -U alpaca_eval
alpaca_eval --model_outputs … --is_recompute_metrics_only True

AlpacaEval 2.0 is a new version of AlpacaEval. Here are the differences:

• reference: gpt4_turbo: we upgraded the baseline from text-davinci-003 to gpt4_turbo to make the benchmark more challenging and have a metric that better reflects the current state of the art.
• annotator: weighted_alpaca_eval_gpt4_turbo: we improved the annotator in quality and price. First, we use the gpt4_turbo model for annotating, which is approximately 2x cheaper than gpt4. Second, we changed the prompt such that the model outputs a single token, which further reduced cost and speed. Finally, instead of using a binary preference, we used the logprobs to compute a continuous preference, which gives the final weighted win-rate. Note that the latter two changes had the surprising effect of decreasing the annotators' length biased.

By default, AlpacaEval 2.0 will be used from pip install alpaca_eval==0.5. If you wish to use the old configs by default, you can set IS_ALPACA_EVAL_2=False in your environment.

As part of AlpacaEval, we release the following data:

• Human annotations (17701) in order to develop and understand automatic evaluators, we release all the human pairwise evaluation that we collected for AlpacaFarm. This contains comparisons between 22 models with the text-davinci-003 reference on the AlpacaFarm evaluation set. Annotations are from a pool of 16 crowd workers on Amazon Mechanical Turk. The different models are: 6 from OpenAI, 2 SFT models from AlpacaFarm, 13 RLHF methods from AlpacaFarm, and LLaMA 7B.
• Human cross-annotations (2596) in order to further analyze automatic evaluators we selected (via stratified sampling across models and datasets) 650 examples from the AlpacaFarm evaluation set and collected 4 human annotations per example.
• AlpacaEval set (805) we made slight modifications/simplification of the AlpacaFarm evaluation set. In particular, we first merged the instruction and input fields into a single instruction field. This affects 1/4 of the examples in the AlpacaFarm evaluation set, all of which are from the self-instruct evaluation set. Second we regenerated the text-davinci-003 reference outputs without limiting the length of its outputs.

For more details about the human annotations refer to the AlpacaFarm paper.

AlpacaEval is an improvement and simplification of the automatic pairwise preference simulator from AlpacaFarm. Outside AlpacaFarm, you should be using AlpacaEval. Here are the main differences:

• AlpacaEval merges instructions and inputs: The AlpacaEval evaluation is the same as the AlpacaFarm evaluation except that the instruction and input fields are merged as {instruction}\n\n{input}. This affects 1/4 of the examples in the AlpacaFarm evaluation set (the self-instruct subset). This simplification provides a more fair comparison for models that were not trained by distinguishing between the two fields.
• AlpacaEval handles longer generations: Models in AlpacaFarm were limited to a maximum number of 300 tokens for generations. We change this number to 2000 for AlpacaEval. Note that this also affects the reference generations (text-davinci-003), so the results on AlpacaEval are not comparable to those on AlpacaFarm even for examples that had no input field.
• AlpacaEval removes intra- and inter-annotator variance: The AlpacaFarm simulator replicates human annotation in terms of both mode behavior and diversity. In particular, AlpacaFarm's simulator uses a pool of models and prompts and adds noise to replicate human intra- and inter-annotator variance. If the goal is to use an automatic annotator for evaluation or simply training better models, then this variance may not be desirable. The default annotators in AlpacaEval thus don't have this variance. We give the option to add it back by using --anotators_config 'alpaca_farm' and --p_label_flip 0.25 when creating an evaluator.

There have been several work that propose new automatic annotators for instruction-following models. Here we list the ones that we are aware of and discuss how they differ from ours. We evaluated all of those in our evaluator's leaderboard.

• Vicuna/lmsys The lmsys annotator (lmsys_gpt4) evaluates the pair by asking the annotator a score from 1-10 for each output, and then selecting the output with the highest score as preferred. They do not randomize over output order and they ask an explanation after the score. Overall, we found that this annotator has strong bias towards longer outputs (0.74) and relatively low correlation with human annotations (63.2).
• AlpacaFarm The best AlpacaFarm annotator (alpaca_farm_greedy_gpt4) evaluates the pair by directly asking the annotator which output it prefers. Furthermore, it batches 5 examples together to amortize the length of the prompt and randomizes the order of outputs. Overall, we found that this annotator has much less bias towards longer outputs (0.60) and is faster (878 seconds/1000 examples) than others. It has a slightly higher correlation with the majority of human annotations (66.4) than humans themselves (65.7). However, it is more expensive ($15.3/1000 examples) and doesn't work with very long outputs given the batching.
• Aviary The Aviary annotator (aviary_gpt4) asks the annotator to order the output by its preference, rather than simply selecting the preferred output. It does not randomize the order of outputs and uses high temperature for decoding (0.9). Overall, we found that this annotator has relatively strong bias towards longer outputs (0.70) and very high correlation with human annotations (69.1). By decreasing the temperature and randomizing the order of outputs, we further improved the correlation to 69.8 (improved_aviary_gpt4) but this further increased the length bias to 0.73.

Our alpaca_eval_gpt4 is a mix between the AlpacaFarm and Aviary annotators. It asks the annotator to order the outputs by preference, but it uses temperature 0, randomizes over outputs, and made some modifications to the prompt to decrease length bias to 0.68.

Other related work include recent papers which analyze automatic evaluators. For example:

• AlpacaFarm Appx C and Large Language Models are not Fair Evaluators both found that automatic annotators have a position bias.
• AlpacaFarm Sec. 5.2. and The False Promise of Imitating Proprietary LLMs both found that automatic annotators favor style (e.g. use of list, tone, word choice, length) over factuality.

For all models you can find the auto-annotations under results/<model_name>/*/annotations.json. The annotations have the following columns:

• instruction: the prompt
• generator_1: the baseline model
• output_1: the output of the baseline model
• generator_2: the model being evaluated
• output_2: the output of the model being evaluated
• annotator: the auto-annotator
• preference: the result of the auto-annotator. This is a float between 1 and 2. Closer to 1 means that the auto-annotator prefers output_1, closer to 2 means that it prefers output_2. For AlpacaEval 2.0, preference-1 corresponds to the probability of output_1 being preferred. For AlpacaEval 1.0, preference is 1 if output_1 is preferred, 2 if output_2 is preferred, and 1.5 if they are the same. The win rate is always(preference -1).mean().
• raw_completion: the raw output of the auto-annotator. This is field contains the completions before de-randomization of the order between output_1 and output_2! It is thus much harder to interpret, see below for more information.

Chain of thought

For some annotators, e.g. alpaca_eval_cot_gpt4_turbo_fn we use chain of thought reasoning to make the models preferences more interpretable. Those can then be found under concise_explanation. To interpret them, you should also look at referenced_models which translates the temporary model name (in the prompt) to the actual output. Below, we provide more explanation as to what is happening behind the scenes.

You can check the raw_annotations["concise_explanation] column in annotations.json (e.g. here) which contains the chain of thought reasoning of the auto annotator. Note that the raw_annotations is not modified by the randomization of the order of the outputs. In particular, "m" and "M" can sometime refer to the first model (the reference) and sometime to the second model (the model being evaluated). To understand which model is being referred to, you should use the column preference and ordered_models. To make it easier we add a column "referenced_models" mapping the model names to the corresponding outputs. For example in the following annotation we see that the preference is 1.0 (i.e. output_1) and corresponds to model M in concise_explanation (see ordered_models).

{
  "instruction": "How did US states get their names?",
  "output_1": "The names of U.S. states are derived from a variety of languages and sources, including indigenous American languages, European languages (primarily English, Spanish, and French), as well as other historical influences. Here is a brief overview of how some states got their names:\n\n1. **Indigenous American Languages**: Many state names come from Native American words or names of tribes, often from the languages of the indigenous peoples who lived in or near those areas.\n   - **Alabama**: Possibly derived from the Choctaw language, meaning \"thicket clearers.\"\n   - **Connecticut**: From a Mohegan-Pequot word meaning \"long tidal river.\"\n   - **Massachusetts**: [...]",
  "generator_1": "gpt4_1106_preview",
  "dataset": "helpful_base",
  "output_2": "The names of the 50 U.S. states come from a variety of sources, including Native American languages, European languages, and historical figures. Here's a brief overview of how some states got their names:\n\n1. Native American origins: Many states have names derived from Native American languages. For example, Alabama comes from the Choctaw word \"Albah amo,\" meaning \"plant gatherers\" or \"herb gatherers.\" Similarly, the name Mississippi comes from the Ojibwe word \"Misi-ziibi,\" meaning \"great river.\"\n\n2. European languages: [...].",
  "generator_2": "gpt4",
  "annotator": "alpaca_eval_cot_gpt4_turbo_fn",
  "preference": 1.0,
  "raw_completion": {
    "concise_explanation": "Model M provided a more detailed and structured response, including bold headings for each category and a wider range of examples. It also included additional categories such as 'Other European Languages' and 'Combination of Languages and Influences', which added depth to the explanation. Model m's response was accurate but less comprehensive and lacked the clear structure found in Model M's output.",
    "ordered_models": [
      {
        "model": "M",
        "rank": 1
      },
      {
        "model": "m",
        "rank": 2
      }
    ]
  },
  "referenced_models": {
    "M": "output_1",
    "m": "output_2"
  }
}

• 12th March 2024: updated to use length-controlled (LC) win rates. This is a debiased version of the win-rates that control for the length of the outputs.
• 3rd January 2024: updated to AlpacaEval 2.0, which uses GPT4-turbo as baseline and annotator.
• 2nd January 2024: added Azure API and more general way of setting client configs. See here
• 19th June 2023: add leaderboard chatgpt_fn that anyone can use (no waiting lists).
• 19th June 2023: update to use OpenAI's function calling. Example: chatgpt_fn or alpaca_eval_gpt4_fn.