feat: AutoFlow on TravelPlanner

.gitignore

@@ -161,5 +161,6 @@ cython_debug/
 log
 openagi_data
 travel_database
+results
 
 .DS_Store

README.md

@@ -33,40 +33,90 @@ conda install pytorch==2.2.2 torchvision==0.17.2 torchaudio==2.2.2 pytorch-cuda=
 
 ```
 pip install -r requirements.txt
+pip install -r requirements_travel.txt
 ```
 
 3. Download the OpenAGI data from this [Google Drive link](https://drive.google.com/drive/folders/1AjT6y7qLIMxcmHhUBG5IE1_5SnCPR57e?usp=share_link), unzip it to the `AutoFlow` directory and rename it as `openagi_data`.
 
-4. Make sure you are in the *AutoFlow/src* folder before running the codes. Otherwise,
+4. Download the [database](https://drive.google.com/file/d/1pF1Sw6pBmq2sFkJvm-LzJOqrmfWoQgxE/view?usp=drive_link) and unzip it to the `AutoFlow` directory (i.e., `your/path/AutoFlow`) and rename it as `travel_database`.
+
+5. Make sure you are in the *AutoFlow/src* folder before running the codes. Otherwise,
 
 ```
 cd src
 ```
 
 ## Running Command Examples
 
-OpenAGI on gpt-4-1106-preview:
+(Notice that --model_name can be different from --auto_model_name)
+
+OpenAGI task when using gpt-4-1106-preview as the workflow interpreter LLM:
 ```commandline
-python auto_agi.py 
+python auto_main.py 
+--flow_name=OpenAGI_Flow.txt
+--task=OpenAGI 
 --model_name="gpt-4-1106-preview"
 --auto_model_name="gpt-4-1106-preview"
 --log_file_name=../log/autoagi_gpt4gpt.txt
 --output_dir=./gpt4gpt
 --auto_flow_name="autoagi_gpt4gpt_Flow.txt"
+--auto_epoch=30 
 --openai_key="YOUR OPENAI KEY"
+--max_round=20
 ```
 
-OpenAGI on TheBloke/Mixtral-8x7B-Instruct-v0.1-GPTQ:
+OpenAGI task when using TheBloke/Mixtral-8x7B-Instruct-v0.1-GPTQ as the workflow interpreter LLM:
 ```commandline
-python auto_agi.py 
+python auto_main.py 
+--flow_name=OpenAGI_Flow.txt
+--task=OpenAGI 
 --model_name="TheBloke/Mixtral-8x7B-Instruct-v0.1-GPTQ"
 --auto_model_name="mistralai/Mixtral-8x7B-Instruct-v0.1"
 --log_file_name=../log/autoagi_mixtral4mixtral.txt
 --output_dir=./mixtral4mixtral
 --auto_flow_name="autoagi_mixtral4mixtral_Flow.txt"
+--auto_epoch=30 
+--openai_key="YOUR OPENAI KEY"
+--max_round=20
+```
+
+TravelPlanner task when using gpt-4-1106-preview as the workflow interpreter LLM:
+```commandline
+python auto_main.py 
+--flow_name=TravelPlanner_Flow_manual.txt 
+--tool_name=tools.txt 
+--task=TravelPlanner 
+--model_name="gpt-4-1106-preview"
+--auto_model_name="gpt-4-1106-preview"
+--log_file_name=../log/auto_travel_gpt4gpt.txt 
+--auto_flow_name=TravelPlanner_Flow_gpt4gpt.txt
+--auto_epoch=30 
 --openai_key="YOUR OPENAI KEY"
+--max_round=100
 ```
 
+TravelPlanner task when using TheBloke/Mixtral-8x7B-Instruct-v0.1-GPTQ as the workflow interpreter LLM:
+```commandline
+python auto_main.py 
+--flow_name=TravelPlanner_Flow_manual.txt 
+--tool_name=tools.txt 
+--task=TravelPlanner 
+--auto_model_name=gpt-4-1106-preview 
+--model_name=TheBloke/Mixtral-8x7B-Instruct-v0.1-GPTQ 
+--log_file_name=../log/auto_travel_gpt4mixtral.txt 
+--auto_flow_name=TravelPlanner_Flow_gpt4mixtral.txt 
+--auto_epoch=30 
+--max_round=100
+```
+
+## Known Issues
+
+[PPOTrainer](https://huggingface.co/docs/trl/main/en/ppo_trainer) class in the latest version of [trl](https://github.com/huggingface/trl) package (==0.9.4) has a known [issue](https://github.com/huggingface/trl/issues/1691) when using multiple GPUs for reinforcement learning.
+
+## Generated Workflow
+
+In the `AutoFlow/src/info/OpenAGI` folder, there are manually designed workflow and automatically generated workflows. The manual workflow is `OpenAGI_Flow_manual.txt`, and the file name of automatically generated workflows is in the form of `OpenAGI_Flow_manual_*4*.txt`. For example, `OpenAGI_Flow_manual_gpt4mixtral.txt` means the workflow is generated by GPT and used for Mixtral as the interpreter LLM.
+
 ## Reference
 
-- We leveraged the dataset of [OpenAGI](https://github.com/agiresearch/OpenAGI) projects and based on [CoRE language] (https://github.com/agiresearch/CoRE) to implement our experiment.
+- We leveraged the dataset of [OpenAGI](https://github.com/agiresearch/OpenAGI) and [TravelPlanner](https://github.com/OSU-NLP-Group/TravelPlanner) projects and based on [CoRE language] (https://github.com/agiresearch/CoRE) to implement our experiment.

requirements_travel.txt

@@ -0,0 +1,9 @@
+langchain==0.1.4
+pandas==2.0.1
+tiktoken==0.4.0
+openai==1.13.3
+langchain_google_genai==0.0.4
+gradio==3.50.2
+datasets==2.15.0
+tiktoken==0.4.0
+func_timeout==4.3.5