smk/graph-of-thoughts
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

- fixed typos and slightly improved language of the README files

Browse Source 
- added README file to the examples directory
- updated the citation section in the main README
This commit is contained in:
Robert Gerstenberger
2023-08-21 11:57:55 +02:00
committed by Nils Blach
parent 41a0a6e67c
commit 2f4b828aef
4 changed files with 23 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files
graph_of_thoughts/operations/README.md
+8 -8
View File
@@ -37,7 +37,7 @@ Remember to set up the predecessors (and optionally successors) for your operati
## Available Operations
The following operations are available in the module:
**Score:** Collect all thoughts from preceeding operations and score them either using the LLM or a custom scoring function.
**Score:** Collect all thoughts from preceding operations and score them either using the LLM or a custom scoring function.
- num_samples (Optional): The number of samples to use for scoring, defaults to 1.
- combined_scoring (Optional): Whether to score all thoughts together in a single prompt or separately, defaults to False.
- scoring_function (Optional): A function that takes in a list of thought states and returns a list of scores for each thought.
@@ -46,25 +46,25 @@ The following operations are available in the module:
- num_samples (Optional): The number of samples to use for validation, defaults to 1.
- improve (Optional): Whether to improve the thought if it is invalid, defaults to True.
- num_tries (Optional): The number of times to try improving the thought, before giving up, defaults to 3.
- validate_funtion (Optional): A function that takes in a thought state and returns a boolean indicating whether the thought is valid.
- validate_function (Optional): A function that takes in a thought state and returns a boolean indicating whether the thought is valid.
**Generate:** Generate new thoughts from the current thoughts. If no previous thoughts are available, the thoughts are initialized with the input to the [Controller](../controller/controller.py).
- num_branches_prompt (Optional): Number of responses that each prompt should generate (passed to prompter). Defaults to 1.
- num_branches_response (Optional): Number of responses the LM should generate for each prompt. Defaults to 1.
- num_branches_response (Optional): Number of responses the LLM should generate for each prompt. Defaults to 1.
**Improve:** Improve the current thoughts. This operation is similar to the ValidateAndImprove operation, but it does not validate the thoughts and always tries to improve them.
**Aggregate:** Aggregate the current thoughts into a single thought. This operation is useful when you want to combine multiple thoughts into a single thought.
- num_responses (Optional): Number of responses to request from the LLM (generates multiple new thoughts). Defaults to 1.
**KeepBestN:** Keep the best N thoughts from the preceeding thoughts. Assumes that the thoughts are already scored and throws an error if they are not.
**KeepBestN:** Keep the best N thoughts from the preceding thoughts. Assumes that the thoughts are already scored and throws an error if they are not.
- n: The number of thoughts to keep in order of score.
- higher_is_better (Optional): Whether higher scores are better (True) or lower scores are better (False). Defaults to True.
**KeepValid:** Keep only the valid thoughts from the preceeding thoughts. Assumes that each thought has already been validated, if not, it will be considered valid.
**KeepValid:** Keep only the valid thoughts from the preceding thoughts. Assumes that each thought has already been validated, if not, it will be considered valid.
**Selector:** Select a number of thoughts from the preceeding thoughts using a selection function. This is useful if subsequent operations should only be applied to a subset of the preceeding thoughts.
**Selector:** Select a number of thoughts from the preceding thoughts using a selection function. This is useful if subsequent operations should only be applied to a subset of the preceding thoughts.
- selector: A function that takes in a list of thoughts and returns a list of thoughts to select.
**GroundTruth**: Evaluates if the preceeding/current thoughts solve the problem and equal the ground truth. This operation is useful for terminating the graph and checking if the final thoughts solve the problem, but is only useful if the ground truth is known.
- ground_truth_evaluator: A function that takes in a thought state and returns a boolean indicating whether the thought solves the problem.
**GroundTruth**: Evaluates if the preceding/current thoughts solve the problem and equal the ground truth. This operation is useful for terminating the graph and checking if the final thoughts solve the problem, but is only useful if the ground truth is known.
- ground_truth_evaluator: A function that takes in a thought state and returns a boolean indicating whether the thought solves the problem.