Create lookup table miners

[steffencruz]

Create lookup table

• Choose a miner model such as pythia and download
• Select a subset of questions from the current Q&A dataset: nq_open. 1000 should be sufficient for now.
• Use the questions as queries for model and store topk responses in a dict: {question: responses}
• Pickle the dict and store in ./lookups/{model_name}_{n_questions}.pkl

Query lookup miner

• Load dict
• When dendrite (miner) is called with a question, use dict as a lookup table and respond with one of the model's responses (using default value if model has not seen that question)

[steffencruz]

then in the config file you can do things like

    dendrite_pool:
      name: DummyDendritePool
      args:
        data_path: nq_open
        miners:
          1: 
            name: pythia-70m
            path: pythia-70m-nq_open-1000.pkl
          2: 
            name: pythia-160m
            path: pythia-160m-nq_open-1000.pkl
          3: 
            name: pythia-70m
            path: pythia-70m-nq_open-1000.pkl
            answer-index: 0
          4: 
            name: pythia-70m
            path: pythia-70m-nq_open-1000.pkl
            answer-index: 0

where the miners specify a path and a name. Note that we can control the diversity of the pool of dendrites by choosing them to be the same or different models, and also you can control which answer they select from the lookup table (answer-index).