Commands
- Running an experiment file
- Running the pipeline
- Run checks on an experiment file
- Create judge file
- Obtain missing results jsonl file
- Convert images to correct form
- Start up Quart server
Running an experiment file
As detailed in the pipeline documentation, you can run a single experiment file using the prompto_run_experiment command and passing in a file. To see all arguments of this command, run prompto_run_experiment --help.
To run a particular experiment file with the data-folder set to the default path ./data, you can use the following command:
This uses the default settings for the pipeline. You can also set the --max-queries, --max-attempts, and --parallel flags as detailed in the pipeline documentation.
If the experiment file is not in the input folder of the data folder, we will make a copy of the file in the input folder which will get processed. If you want to move the file to the input folder, you can use the --move-to-input flag:
Note that if the experiment file is already in the input folder, we will not make a copy of the file and process the file in place.
Automatic evaluation using an LLM-as-judge
It is possible to automatically run a LLM-as-judge evaluation of the responses by using the --judge-folder and --judge arguments of the CLI. See the Create judge file section for more details on these arguments.
For instance, to run an experiment file with automatic evaluation using a judge, you can use the following command:
prompto_run_experiment \
--file path/to/experiment.jsonl \
--data-folder data \
--judge-folder judge \
--judge gemini-1.0-pro
Running the pipeline
As detailed in the pipeline documentation, you can run the pipeline using the prompto_run_pipeline command. To see all arguments of this command, run prompto_run_pipeline --help.
To run a particular experiment file with the data-folder set to pipeline-data, you can use the following command:
This uses the default settings for the pipeline. You can also set the --max-queries, --max-attempts, and --parallel flags as detailed in the pipeline documentation.
Run checks on an experiment file
It is possible to run a check over an experiment file to ensure that all the prompts are valid and the experiment file is correctly formatted. We also check for environment variables and log any errors or warnings that are found. To run this check, you can use the prompto_check_experiment command and passing in a file. To see all arguments of this command, run prompto_check_experiment --help.
To run a check on a particular experiment file, you can use the following command:
This will run the checks on the experiment file and log any errors or warnings that are found. You can optionally set the log-file to save the logs to a file using the --log-file flag (by default, it will be saved to a file in the current directory) and specify the path to the data folder using the --data-folder flag.
Lastly, it’s possible to automatically move the file to the input folder of the data folder if it is not already there. To do this, you can use the --move-to-input flag:
prompto_check_experiment \
--file path/to/experiment.jsonl \
--data-folder data \
--log-file path/to/logfile.txt \
--move-to-input
Create judge file
Once an experiment has been ran and responses to prompts have been obtained, it is possible to use another LLM as a “judge” to score the responses. This is useful for evaluating the quality of the responses obtained from the model. To create a judge file, you can use the prompto_create_judge_file command passing in the file containing the completed experiment and to a folder (i.e. judge folder) containing the judge template and settings to use. To see all arguments of this command, run prompto_create_judge_file --help.
To create a judge file for a particular experiment file with a judge-folder as ./judge and using judge gemini-1.0-pro you can use the following command:
prompto_create_judge_file \
--experiment-file path/to/experiment.jsonl \
--judge-folder judge \
--templates template.txt \
--judge gemini-1.0-pro
In judge, you must have the following files:
settings.json: this is the settings json file which contains the settings for the judge(s). The keys are judge identifiers and the values dictionaries with “api”, “model_name”, “parameters” keys to specify the LLM to use as a judge (see the experiment file documentation for more details on these keys).- template
.txtfile(s) which specifies the template to use for the judge. The inputs and outputs of the completed experiment file are used to generate the prompts for the judge. This file should contain the placeholders{INPUT_PROMPT}and{OUTPUT_RESPONSE}which will be replaced with the inputs and outputs of the completed experiment file (i.e. the corresponding values to thepromptandresponsekeys in the prompt dictionaries of the completed experiment file).
For the template file(s), we allow for specifying multiple templates (for different evaluation prompts), in which case the --templates argument should be a comma-separated list of template files. By default, this is set to template.txt if not specified. In the above example, we explicitly pass in template.txt to the --templates argument, so the command will look for a template.txt file in the judge folder.
See for example this judge example which contains example template and settings files.
The judge specified with the --judge flag should be a key in the settings.json file in the judge folder. You can create different judge files using different LLMs as judge by specifying a different judge identifier from the keys in the settings.json file.
Obtain missing results jsonl file
In some cases, you may have ran an experiment file and obtained responses for some prompts but not all (e.g. in the case where an experiment was stopped during the process). To obtain the missing results jsonl file, you can use the prompto_obtain_missing_results command passing in the input experiment file and the corresponding output experiment. You must also specify a path to a new jsonl file which will be created if any prompts are missing in the output file. The command looks at an ID key in the prompt_dicts of the input and output files to match the prompts, by default the name of this key is id. If the key is different, you can specify it using the --id flag. To see all arguments of this command, run prompto_obtain_missing_results --help.
To obtain the missing results jsonl file for a particular experiment file with the input experiment file as path/to/experiment.jsonl, the output experiment file as path/to/experiment-output.jsonl, and the new jsonl file as path/to/missing-results.jsonl, you can use the following command:
prompto_obtain_missing_results \
--input-experiment path/to/experiment.jsonl \
--output-experiment path/to/experiment-output.jsonl \
--missing-results path/to/missing-results.jsonl
Convert images to correct form
The prompto_convert_images command can be used to convert images to the correct form for the multimodal LLMs. This command takes in a folder containing images and checks if .jpg, .jpeg and .png files are saved in the correct format. If not, we resave them in the correct format.
To convert images in a folder ./images to the correct form, you can use the following command:
Start up Quart server
As described in the Quart API model documentation, we have implemented a simple script to start up a Quart API that can be used to query a text-generation model from the Huggingface model hub using the Huggingface transformers library. To start up the Quart server, you can use the prompto_start_quart_server command along with the Huggingface model name. To see all arguments of this command, run prompto_start_quart_server --help.
To start up the Quart server with vicgalle/gpt2-open-instruct-v1, at "http://localhost:8000", you can use the following command: