Support combining multiple columns into a single text feature

[tgaddair]

For LLMs and other text models, it's common to treat multiple inputs as a single string of text to feed into the model as input, rather than construct separate "towers" for each feature. This PR provides a standard way to do this via the prompt config.

For ECD:

input_features:
    - name: context
      type: text
      preprocessing:
          prompt:
            template: The {color} {animal} jumped over the {size} {object}

For LLM:

model_type: llm
prompt:
    template: The {color} {animal} jumped over the {size} {object}
input_features:
    - name: context
      type: text

Here, it is not required that context be a valid column of the input dataset, as it will be generated from the prompt. However, each of color, animal, size, and object are assumed to be columns of the input dataset.

[github-actions]

Unit Test Results

  6 files  ±0    6 suites  ±0   1h 22m 39s ⏱️ + 5m 23s
33 tests ±0  29 ✔️ ±0    4 💤 ±0  0 ❌ ±0 
99 runs  ±0  87 ✔️ ±0  12 💤 ±0  0 ❌ ±0 

Results for commit 16ebf37. ± Comparison against base commit 2bada30.

♻️ This comment has been updated with latest results.

[justinxzhao]

LGTM! Just 2 small comments.

[justinxzhao]

OOC: Would a user be able to use the DEFAULT_ZERO_SHOT_PROMPT_TEMPLATE for the prompt template within an input feature definition like so?

input_features:
    - name: context
      type: text
      preprocessing:
          prompt:
            template: None  # Use default.

[tgaddair]

Yes, though I believe in this case they need to provide a task.

[arnavgarg1]

Yeah that's right, they'd need to provide a task

[geoffreyangus]

How does this work? Are we accepting prompt keys at both the local and global level?

[tgaddair]

The LLM schema requires the prompt at the top-level, and does not allow feature-level prompts.

The ECD schema allows prompts at the feature-level, but not the top level.

[geoffreyangus]

why do we need to cast the values?

[tgaddair]

If you have a number like 0.1234 in the DF and you want to format it like {number:.2f} it will fail if the number is represented as a string in the DF. So we need to cast it to a float. I can add a comment to this effect.