Add option to outlines adm to filter votes to positive responses only

CHANGELOG.md

@@ -28,6 +28,7 @@ This project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.htm
 * Added dedicated function to utils for calculating votes (same voting scheme as the single KDMA ADM)
 * Added top level config options to force determinism and fix seeds; along with an example experiment to demonstrate
 * Added sampler parameter to outlines ADMs (example usage in `align_system/configs/experiment/examples/outlines_sampler.yaml`)
+* Added option (on by default) to outlines ADM to filter votes to positive options only, can disable on the command line with `+adm.inference_kwargs.filter_votes_to_positives=False`
 
 ### Deprecated 
 * The algorithm `align_system/algorithms/chat_kdma_predicting_adm.py` has been replaced by `align_system/algorithms/outlines_regression_adm.py`

align_system/algorithms/outlines_adm.py

@@ -15,7 +15,10 @@
 
 from align_system.utils import logging
 from align_system.utils import get_swagger_class_enum_values
-from align_system.utils.voting import calculate_votes
+from align_system.utils.voting import (
+    calculate_votes,
+    filter_votes_to_responses,
+)
 from align_system.algorithms.abstracts import ActionBasedADM
 from align_system.prompt_engineering.outlines_prompts import (
     baseline_system_prompt,
@@ -221,8 +224,22 @@ def top_level_choose_action(self,
                     extra={"markup": True})
         log.explain(votes, extra={"highlighter": JSON_HIGHLIGHTER})
 
+        if kwargs.get('filter_votes_to_positives', True):
+            filtered_votes = filter_votes_to_responses(
+                votes, positive_responses_choices)
+
+            if filtered_votes != votes:
+                log.explain("Filtering votes down to choices where we "
+                            "have a positive reponse")
+                log.explain(filtered_votes,
+                            extra={"highlighter": JSON_HIGHLIGHTER})
+
+            final_votes = filtered_votes
+        else:
+            final_votes = votes
+
         # Take top choice by score (votes is a dictionary of choice: score)
-        top_choice, top_choice_score = max(votes.items(), key=lambda x: x[1])
+        top_choice, top_choice_score = max(final_votes.items(), key=lambda x: x[1])
         # Just taking first justification from the positive responses
         # where the top choice was selected.  A better approach might
         # be to somehow summarized all justifications with the

align_system/utils/voting.py

@@ -33,3 +33,14 @@ def calculate_votes(possible_choices,
                         for choice, score in tmp_normalized_votes.items()}
 
     return normalized_votes
+
+
+def filter_votes_to_responses(votes, responses):
+    filtered_votes = {choice: score for choice, score in votes.items()
+                      if choice in responses}
+
+    if len(filtered_votes) == 0:
+        raise RuntimeError(
+            "No votes left after filtering, was `reponses` empty?")
+
+    return filtered_votes