Using the library with any open source llm

[harsh3986]

How can I use pandas ai with any open source llms or let's say any model from hugging face ?

[sandiemann]

this doc might be out to date but you can initiate the HF llm models Starcoder or Falcon as mentioned here

[harsh3986]

Thanks @sandiemann. Suppose I have created a fine tuned starcoder model, how will I use it in that case ?

[sandiemann]

sample code:

from pandasai import PandasAI
from pandasai.llm.starcoder import Starcoder

df = pd.read_csv("examples/data/Loan payments data.csv")

#llm = OpenAI()
llm = Starcoder()
pandas_ai = PandasAI(llm, verbose=True, conversational=True)
response = pandas_ai(df, "How many loans are from men and have been paid off?")
print(response)

output logs:

2023-07-18 23:51:05 [INFO] Question: How many loans are from men and have been paid off?
2023-07-18 23:51:05 [INFO] Running PandasAI with starcoder LLM...
2023-07-18 23:51:05 [INFO] Prompt ID: 392db83b-6b36-4d71-98f7-1b5ab0bfa406
2023-07-18 23:51:07 [INFO] 
                        Code generated:
                        ```
                        df[df['Gender']=='male']
                        df[df['Gender']=='male']['loan_status'].value_counts()
                        ```
                    
2023-07-18 23:51:07 [INFO] 
Code running:

df[df['Gender'] == 'male']
df[df['Gender'] == 'male']['loan_status'].value_counts()

2023-07-18 23:51:07 [INFO] Answer: PAIDOFF               247
COLLECTION             90
COLLECTION_PAIDOFF     86
Name: loan_status, dtype: int64

[mspronesti]

@harsh3986 In the case of a fine-tuned LLM, PandasAI does not natively support any solution as it relies on the hosted inference api for hf models. However, since it is integrated with langchain, you might try to load your fine-tuned StarCoder as a langchain model (e.g. assuming you uploaded it to huggingface hub, you can use this model) and pass it to the PandasAI class, which internally converts it.
Hope this helps.