Skip to main content

AlpacaEval : An Automatic Evaluator of Instruction-following Models

Project description

AlpacaEval : An Automatic Evaluator for Instruction-following Language Models

Code License Data License Python 3.10+ discord

Evaluation of instruction-following models (e.g., ChatGPT) typically requires human interactions. This is time-consuming, expensive, and hard to replicate. AlpacaEval in an LLM-based automatic evaluation that is fast, cheap, replicable, and validated against 20K human annotations. It is particularly useful for model development. Although we improved over prior automatic evaluation pipelines, there are still fundamental limitations. AlpacaEval provides the following:

  • Automatic evaluator: an automatic evaluator that has high agreement with humans (validated on 20K annotations). We evaluate a model by measuring the fraction of times an powerful LLM (e.g. GPT 4 or Claude) prefers the outputs from that model over outputs from a reference model. Our evaluators enable caching and output randomization by default.
  • Leaderboard: a leaderboard of common models on the AlpacaEval evaluation set.
  • Toolkit for building automatic evaluators: a simple interface for building advanced automatic evaluators (e.g. with caching, batching, or multi-annotators) and analyzing them (quality, price, speed, statistical power, bias, variance etc).
  • Human evaluation data: 20K human preferences between a given and reference model on the AlpacaFarm evaluation set. 2.5K of these are cross-annotations (4 humans annotating the same 650 examples).
  • AlpacaEval dataset: a simplification of AlpacaFarm's evaluation set, where "instructions" and " inputs" are merged into one field, and reference outputs are longer.

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.

Table of Contents
  1. Quick Start
  2. Leaderboards and how to interpret them
  3. Use-cases
  4. Analysis
  5. Contributing
  6. Limitations
  7. Citation
  8. Additional information

Quick Start

To install the stable release, run

pip install alpaca-eval

To install the nightly version, run

pip install git+https://github.com/tatsu-lab/alpaca_eval

Then you can use it as follows:

export OPENAI_API_KEY=<your_api_key>
export OPENAI_ORGANIZATION_IDS=<your_organization_id>  # Optional; if not set, this will be your default org id.
alpaca_eval --model_outputs 'example/outputs.json' 

Important parameters are the following:

  • model_outputs : A path to a json file for the outputs of the model to add to the leaderboard. Each dictionary should contain the keys instruction and output.
  • annotators_config: This is the annotator to use (e.g., alpaca_eval_gpt4 or claude). alpaca_eval_gpt4 ( default) has the highest agreement rate with our human annotation data. claude has a decent agreement and is free for academics. For a comparison of annotators see here.
  • reference_outputs: The outputs of the reference model. Same format as model_outputs. By default, this is text-davinci003 outputs on AlpacaEval dataset.
  • output_path: Path for saving annotations and leaderboard.

If you don't have the model outputs, you can use evaluate_from_model and pass a local path or a name of a HuggingFace model, or a model from a standard API (OpenAI, Anthropic, Cohere). Other commands:

>>> alpaca_eval -- --help
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 (agreement with human, speed, price,...).

For more information about each function use alpaca_eval <command> -- --help.

Leaderboards and how to interpret them

Models

Our leaderboards are computed on the AlpacaEval dataset. We precomputed the leaderboard for important models both using gpt4 (best quality) and claude (free for academics, and high quality). Our full leaderboards can be found at on this page, but we give minimal leaderboards below. Later we also show how to add your model to the leaderboard and how to make a new leaderboard for your evaluator/dataset. See here for the configs of all models that are available out of the box.

alpaca_eval_gpt4 minimal leaderboard:

Win Rate Std Error
gpt4 95.3 0.7
claude 88.4 1.1
chatgpt 86.1 1.2
wizardlm-13b 75.3 1.5
guanaco-65b 71.8 1.6
vicuna-13b 70.4 1.6
oasst-rlhf-llama-33b 66.5 1.7
text_davinci_003 50.0 0.0
falcon-40b-instruct 45.7 1.8
alpaca-farm-ppo-human 41.2 1.7
alpaca-7b 26.5 1.5
text_davinci_001 15.2 1.2
How exactly are those metrics computed?

Win Rate: the win rate measures the fraction of time the model's output is preferred over text-davinci-003 outputs ( i.e. the reference). 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 (text-davinci-003) on the same instruction. We then ask our automatic evaluator which output they prefer. See here and here for the exact prompts and configs for GPT4 and Claude, 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 text-davinci-003. 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.

Details about alpaca_eval_gpt4

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.

claude minimal leaderboard
Win Rate Std Error
gpt4 77.0 1.5
claude 75.8 1.5
chatgpt 67.7 1.6
wizardlm-13b 66.1 1.7
vicuna-13b 63.2 1.7
guanaco-65b 62.6 1.7
oasst-rlhf-llama-33b 57.3 1.7
text_davinci_003 50.0 0.0
falcon-40b-instruct 46.7 1.8
alpaca-farm-ppo-human 46.5 1.8
alpaca-7b 32.3 1.6
text_davinci_001 21.5 1.4

Evaluators

We evaluate different automatic annotators on the AlpacaEval set by comparing to 2.5K human annotations we collected (~650 instructions each with 4 human annotations). Below we show metrics for our suggested evaluator (alpaca_eval_gpt4), for prior automatic evaluators (alpaca_farm_greedy_gpt4,aviary_gpt4,lmsys_gpt4), for humans (humans), and for different base models with essentially the same prompt (gpt4,claude,text_davinci_003,guanaco_33b, chatgpt). See here for the configs of all evaluators that are available out of the box and their associated metrics.

Human agreement [%] Price [$/1000 examples] Time [seconds/1000 examples] Bias Variance Proba. prefer longer
alpaca_eval_gpt4 69.2 13.6 1455 28.4 14.6 0.68
aviary_gpt4 69.1 12.8 1869 29.5 13.1 0.70
gpt4 66.9 12.5 1037 31.5 14.6 0.65
alpaca_farm_greedy_gpt4 66.4 15.3 878 30.2 19.3 0.60
humans 65.7 300.0 36800 0.0 0.64
claude 65.5 11.1 173 31.9 18.0 0.62
text_davinci_003 64.1 8.7 121 33.8 22.7 0.70
lmsys_gpt4 63.2 13.9 17982 34.7 16.1 0.74
guanaco_33b 59.1 930 54.5 27.1 0.70
chatgpt 57.2 0.8 285 39.4 34.1 0.59
How exactly are those metrics computed?

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 ($18/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.

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.

Tips for choosing evaluators

Overall we recommend using annotators_config=alpaca_eval_gpt4 if you want the highest agreement with humans, and annotators_config=claude if you have academic (free) access to Claude and have a low 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]"
  • "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 alpaca_eval_gpt4 to be a good trade-off between quality / price / time / variance / length bias.

The above metrics are computed with respect to annotations from crowd-workers. Although useful, those annotations are not perfect, e.g., crowd-workers often favor style over factuality. We thus recommend users to validate automatic evaluators on their own instructions and human annotations. Details in limitations.

Use-cases

Evaluating a model

>>> alpaca_eval evaluate -- --help
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
        Defaul...
        The outputs of the reference model. Same format as `model_outputs`. If None, the reference outputs are the
 003 outputs on the AlpacaEval set.
    --annotators_config=ANNOTATORS_CONFIG
        Type: Union
        Default: 'alpaca_eval_gpt4'
        The path the (or list of dict of) the annotator's config file. For details see the docstring of `PairwiseA
nnotator`.
    -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 no
t 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 sav
e. 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 t
he column `win_rate`. If `auto` we will try to use the corresponding leaderboard for the reference outputs (only i
f 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 func
tion should take a sequence of preferences (0 for draw, 1 for base win, 2 when the model to compare wins) and retu
rn 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
        Default: False
        Whether to save the result leaderboard to `precomputed_leaderboard`. If None we save only if max_instances
. 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`.
    Additional flags are accepted.
        Additional arguments to pass to `PairwiseAnnotator`.
>>> alpaca_eval evaluate_from_model -- --help
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 same model as the one we are
    -e, --evaluation_dataset=EVALUATION_DATASET
        Type: Union
        Defaul...
        Path to the evaluation dataset or a function that returns a dataframe. If None, we use the default evaluat
ion
    -a, --annotators_config=ANNOTATORS_CONFIG
        Type: Union
        Default: 'alpaca_eval_gpt4'
        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.
    -i, --is_strip_output=IS_STRIP_OUTPUT
        Type: bool
        Default: True
        Whether to strip trailing and leading whitespaces from the outputs.
    Additional flags are accepted.
        Other kwargs to `evaluate`

NOTES
    You can also use flags syntax for POSITIONAL ARGUMENTS

To evaluate a model you need to:

  1. Choose an evaluation set and compute outputs specified as model_outputs. By default, we use the 805 examples from AlpacaEval. To compute outputs on AlpacaEval use:
import datasets

eval_set = datasets.load_dataset("tatsu-lab/alpaca_eval", "alpaca_eval")["eval"]
for example in eval_set:
    # generate here is a placeholder for your models generations
    example["output"] = generate(example["instruction"])

if your model is a HuggingFace model or from a standard API provider (OpenAI, Anthropic, Cohere). Then you can directly use alpaca_eval evaluate_from_model to also take care of generating outputs.

  1. Compute the reference outputs reference_outputs. By default, we use the outputs of text-davinci-003 on AlpacaEval. If you want to use a different model or a different dataset follow the same steps as (1.).
  2. Choose an evaluator specified via annotators_config. We recommend using alpaca_eval_gpt4 or claude (if you are an academic). For options and comparisons see this table. Depending on the evaluator you might need to set the appropriate API_KEY in your environment or here.

Running all together:

alpaca_eval --model_outputs 'example/outputs.json' \
  --annotators_config 'alpaca_eval_gpt4' \
  --reference_outputs <path to outputs if not text_davinci_003 on AlpacaEval>

If you don't have decoded outputs, you can use evaluate_from_model which takes care of decoding (model and reference) for you. Here's an example:

# need a GPU for local models
export ANTHROPIC_API_KEY=<your_api_key> # let's annotate with claude
alpaca_eval evaluate_from_model \
  --model_configs 'oasst_pythia_12b' \
  --annotators_config 'claude' \
  --reference_model_configs <path to configs not text_davinci_003 on AlpacaEval>        

Here the model_configs and reference_model_configs (optional) are paths to a directory that specifies the prompt, the model provider (here HuggingFace) and decoding parameters. See this directory for examples. For all model providers that are available out-of-the-box see here.

Information about annotators
  • 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.

Making a new leaderboard

>>> alpaca_eval make_leaderboard -- --help
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 LEADERBOARD_PATH <flags>

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

POSITIONAL ARGUMENTS
    LEADERBOARD_PATH
        Type: Union
        The path to save the leaderboard to. The leaderboard will be saved as a csv file, if it already exists it will

FLAGS
    --annotators_config=ANNOTATORS_CONFIG
        Type: Union
        Default: 'alpaca_eval_gpt4'
        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, datas
ets.Dataset) or a path to read those (json, csv, tsv potentially with globbing) or a function to generate those. I
f the path contains a globbing pattern, we will read all files matching the pattern and concatenate them. Each dic
tionary (or row of dataframe) should contain the keys that are formatted in the prompts. E.g. by default `instruct
ion` 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
        Defaul...
        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_lea
derboard`, `is_return_instead_of_print`, `reference_outputs`.
    -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`.

NOTES
    You can also use flags syntax for POSITIONAL ARGUMENTS

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.

Making a new evaluator

>>> alpaca_eval analyze_evaluators -- --help
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 (agreement with human, speed, price,...).

FLAGS
    --annotators_config=ANNOTATORS_CONFIG
        Type: Union
        Default: 'alpaca_eval_gpt4'
        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[]
        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.

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].
Other parameters in the configuration file

The easiest is to check the docstrings of SinglePairwiseAnnotator. Here are some important ones:

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.

Additional analysis and plots

AlpacaEval provides a few visualization tools to help you analyze and improve your automatic evaluation pipeline. We briefly explain them here and provide notebooks for more analysis. For a description of all the metrics we consider refer to How exactly are those metrics computed?

Analyzing an evaluator

Analyzing evaluators: analyzing an evaluator

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.

plot_quality_vs_price_and_time.png

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 claude evaluator.

Correlation between humans and alpaca_eval_gpt4

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.

Analyzing an eval set

Making evaluation sets: analyzing an evaluator

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.

Number of samples needed to distinguish pairs in the Claude leaderboard

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.

plot_paired_ttests_per_dataset.png

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.

Contributing

We are accepting PRs for new models, evaluators, and eval sets, in addition to bug fixes. We will update the leaderboard website regularly with new community contributions. We have also created a support discord for AlpacaEval in case you run into any issues and wish to ask help from the community.

Contributing a model

First, you'll need to add a model config definition in the models_configs folder. As an example, you can look at the falcon-7b-instruct yaml. Please make sure the folder name and key name in the yaml match exactly.

Then, please follow the steps in Evaluating-a-model to run inference on the model to produce outputs on the eval set and score the model according to one of the evaluators. An example command may look like:

alpaca_eval evaluate_from_model \
  --model_configs 'falcon-7b-instruct' \
  --annotators_config 'alpaca_eval_gpt4'

After running this command, you should have generated an outputs json and a new entry in the corresponding leaderboard file. Please make a PR with the config, outputs file, and updated leaderboard.

Contributing an evaluator

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/your_leaderboard_name.csv> \
  --all_model_outputs alpaca_eval_all_outputs.json \
  --annotators_config <path_to_your_config.yaml>

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

Contributing an eval set

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.

Limitations

The AlpacaEval evaluation pipeline, like other current evaluators have important limitations and should therefore not be used as replacement for human evaluation in important settings, such as to decide whether a model is ready to be deployed. Those can broadly be clustered into 3 categories:

  1. Instructions might not be representative of real-usage: the AlpacaEval set contains examples from a variety of datasets (self-instruct, open-assistant, vicuna, koala, hh-rlhf) which might not be representative of real-usage and advanced applications of better models like GPT4. As a result, the gap between the top and the rest of the AlpacaEval leaderboard is likely smaller than it would be on more complex instructions. See for example this blog for preliminary results on more complex instructions. Note, however, that in AlpacaFarm we showed that win-rates on our evaluation set are highly correlated (0.97 R2) with win-rates on instructions from user interactions with the Alpaca Demo. Furthermore, the AlpacaEval leaderboard shows larger gap between the open models and OpenAI models than other leaderboards ( e.g. lmsys).

  2. Biases of automatic annotators: the automatic annotators seem to have implicit biases. In particular, we found that they tend to prefer longer outputs and outputs that contain lists (e.g. 0.68 / 0.69 for alpaca_eval_gpt4 and 0.62 / 0.58 for claude). Although we found that humans have similar biases (0.64 / 0.61), we believe that this could be more of a limitation of human annotation pipeline we used rather than a true human bias. More generally, through qualitative analysis, we found that automatic annotators give more importance to the style of the output than its content (e.g. factuality). Finally, we found that automatic evaluators tend to prefer outputs from models that are similar (likely trained on the same data) as suggested by the big difference between ChatGPT/GPT4 on claude's and alpaca_eval_gpt4's leaderboard.

  3. Lack of safety evaluation: importantly, AlpacaEval only evaluates the instruction-following capabilities of models rather than the harm that they could cause (e.g. toxic behavior or bias). As a result the small gap between current ChatGPT and the best open source models should not be interpreted as if that the latter are ready to be deployed.

Beyond those limitations about the evaluation pipelines, there are also limitations about our validation of the evaluators and our proposed approach to selecting evaluation sets.

Limitations about our validation pipeline

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.

Citation

Please consider citing the repo if you used the automatic annotators, code, or results.

@misc{alpaca_eval,
  author = {Xuechen Li and Tianyi Zhang and Yann Dubois and Rohan Taori and Ishaan Gulrajani and Carlos Guestrin and Percy Liang and Tatsunori B. Hashimoto },
  title = {AlpacaEval: An Automatic Evaluator of Instruction-following Models},
  year = {2023},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/tatsu-lab/alpaca_eval}}
}

If you used our human annotation data, please also consider citing the AlpacaFarm paper:

@misc{dubois2023alpacafarm,
  title={AlpacaFarm: A Simulation Framework for Methods that Learn from Human Feedback}, 
  author={Yann Dubois and Xuechen Li and Rohan Taori and Tianyi Zhang and Ishaan Gulrajani and Jimmy Ba and Carlos Guestrin and Percy Liang and Tatsunori B. Hashimoto},
  year={2023},
  eprint={2305.14387},
  archivePrefix={arXiv},
  primaryClass={cs.LG}
}

If you use the AlpacaEval evaluation set, please cite each of the constituent datasets: self-instruct, open-assistant, vicuna, koala, hh-rlhf.

More information

Data Release

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.

Differences with AlpacaFarm

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.

Related work

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:

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

alpaca_eval-0.1.3.tar.gz (274.5 kB view hashes)

Uploaded Source

Built Distribution

alpaca_eval-0.1.3-py3-none-any.whl (127.4 kB view hashes)

Uploaded Python 3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page