ianwu13/llm_nego_plan_exps

LLM Negotiation Planning Experiments

Contains code (currently just structure) for experiments on incorporating dialogue planning into LLMs, specifically in the domain of resource negotiation.

Directory Structure - UPDATE

• agent_experiments/ : base directory for experiments

  • agents/ : code/classes for different types of agents, level of interaction (utterance/act, planned/not planned) is handled here

    • plan_to_utt_agent.py : generates utterance responses based on dialogue act level interactions
    • planning_agent.py : generates dialogue act responses based on dialogue act history
    • utterance_agent.py : generates utterance responses based on utterance history

  • data/ : scripts to return data from different formats/datasets in a consistent format

    • conversion/ : code to convert/annotate existing datasets to alternative formats
    • datasets/ : raw files contiaining datasets

  • interactions/ : code to allow for interaction with/between different models

  • llm_apis/ : scripts to allow for interfacing with diffferent LLM APIs that are hosted elsewhere

  • misc/ : miscellaneous one-off scripts for any tasks related to the project - as required.

  • models/ : weights for dialogue planning models

  • do_chat.py

  • do_selfplay.py

  • commd_gen_selfplay.py: script for generating commd of running different experiments and stored as .json and .md file

  • extract.py: Recursively extract dialogue content from every log file into json format in the provided input directory and stored the result to provided output directory

  • get_deal.py: detect the final deal from conversation in json format

  • get_prompts.py:

  • registry.py: A registry for supported (dataset, model, task) triplets and required name to class mappings.

  • utils.py: Independent utilities used throughout the code.

  • Makefile: Automaticly running all the experiment commd in commd.json file

  • storage/ : SUGGESTED - for all storage, gitignored by default

    • logs/ : log dir for the results and outputs of the tasks.
    • utilities/ : additional stuff like commands and API keys for connecting to the models. - should never be pushed to the code repository.

Data Annotation

• Data annotation can be run using the annotate/py script. This script utilizes LLMs to annotate data. The LLM and prompt to use are specified as commnad line arguments when calling the script. Pre-implemented LLM classes and prompts with their respective argument values can be found in registry.py.

Adding a Dataset

• Pipelines have been implemented for the CaSiNo and DND datasets. Datasets are implemented as subclasses of the BaseDatasetHandler class found in agent_experiments/data/datahandler.py. To add a dataset, simply extend this class and its methods for the new datasets, and add the new class to registry.py. Examples of this can be seen in agent_experiments/data/casino.py and agent_experiments/data/dealornodeal.py

Adding a Data Annotation pipeline

• Annotation pipelines are specified by the command line arguments provided to annotate.py Beyond the LLM and dataset, the most important part is the "--inst_to_prompt_funct" argument, which specifies the function that converts a data point into a prompt to feed the LLM. Examples of these funcitons can be found in agent_experiments/data/conversion/inst2p_functions.py.

Selfplay

• Check the required files and experiment agents parameters in commd_gen_selfplay.py and get the experiment command file by "running: python commd_gen_selfplay.py dataset_name api_key" It will generate two files for each dataset, containing the command for running the selfplay experiment. commd_selfplay_dataset.json (file for makefile to read and run) and commd_selfplay_dataset.md (for human to check command)

• Running all experiment of a dataset by "make dataset_name". (e.g. make casino) The experiment result will be stored as log file under the provided output directory in commands.

• Extract conversation content from log file and store as json format by running "python extract.py dataset". Make sure the input and output directory is valid.

• Get the final deal from GPT by providing the conversation contents, running "python get_deal.py dataset api_key" will append the final deal to the input json file. It will also print the basic stats categorizing different deals and stored as seperate file for each agent_agent scenarios.

Chat

• bot v. bot chat cna be run using the do_chat.py script. Arguments to this script are essentially identical to those for do_selfplay.py, but can be specified seperately for each of the bots acting in the negotiations.

---

Acknowledgements

https://arxiv.org/abs/1512.03385

https://github.com/facebookresearch/end-to-end-negotiator

---

Adding a new dataset, model, or task

1. Update the registry variables.
2. Follow existing templates in the codebase to add a new class for the dataset, model, or task.

Formatting Guidelines

• Keep this README up-to-date.
• Use classes and inheritence - as required.
• Follow the naming conventions and line spacing properly - 2 lines for the items in global scope, 1 line for the items within a local scope.
• Add comments properly and add citations to external sources
• Add links in the comments if you copy code directly from any external resource.
• Avoid redundant code - use utils.py to put code that can be shared between multiple classes. Use class inheritence to share any common code.
• Use uniformity (very important)