SLIM-QA-GEN-PHI-3-TOOL

slim-qa-gen-phi-3-tool is a 4_K_M quantized GGUF version of slim-qa-gen-phi-3, providing a small, fast inference implementation, optimized for multi-model concurrent deployment.

This model implements a generative 'question' and 'answer' (e.g., 'qa-gen') function, which takes a context passage as an input, and then generates as an output a python dictionary consisting of two keys:

 `{'question': ['What was the amount of revenue in the quarter?'], 'answer': ['$3.2 billion']} `  

The model has been designed to accept one of three different parameters to guide the type of question-answer created:
-- 'question, answer' (generates a standard question and answer),
-- 'boolean' (generates a 'yes-no' question and answer), and
-- 'multiple choice' (generates a multiple choice question and answer).

Note: we would generally recommend using sampling and temperature(0.5+) for varied generations, although if using 'multiple choice' mode, then we have seen the best results with temperature in the 0.2-0.3 range.

slim-qa-gen-phi-3 is the Pytorch version of the model, and suitable for fine-tuning for further domain adaptation.

To pull the model via API:

from huggingface_hub import snapshot_download           
snapshot_download("llmware/slim-qa-gen-phi-3-tool", local_dir="/path/on/your/machine/", local_dir_use_symlinks=False)  

Load in your favorite GGUF inference engine, or try with llmware as follows:

from llmware.models import ModelCatalog  

# to load the model and make a basic inference
model = ModelCatalog().load_model("slim-qa-gen-phi-3-tool", temperature=0.5, sample=True)  
response = model.function_call(text_sample, params=["boolean"])  

# this one line will download the model and run a series of tests  
ModelCatalog().tool_test_run("slim-qa-gen-phi-3-tool", verbose=True)    

Note: please review config.json in the repository for prompt template information, details on the model, and full test set.

Model Card Contact

Darren Oberst & llmware team

Any questions? Join us on Discord