Improve whodunnit

PiperOrigin-RevId: 633613807 Change-Id: I9b3223ce963f9ce2e647bc93ae9a985023a51967
google-deepmind · May 14, 2024 · 79c2043 · 79c2043
1 parent b5a0be1
commit 79c2043
Show file tree

Hide file tree

Showing 3 changed files with 29 additions and 9 deletions.
diff --git a/concordia/components/agent/justify_recent_voluntary_actions.py b/concordia/components/agent/justify_recent_voluntary_actions.py
@@ -134,6 +134,7 @@ def update(self) -> None:
         ),
         max_characters=2000,
         max_tokens=1000,
+        terminators=(),
     )
     what_effect_it_had = what_they_did_chain_of_thought.open_question(
         question=(
@@ -143,6 +144,7 @@ def update(self) -> None:
         ),
         max_characters=2000,
         max_tokens=1000,
+        terminators=(),
     )
     # Now consider how to justify the voluntary actions for all audiences.
     justification_chain_of_thought = interactive_document.InteractiveDocument(
@@ -169,6 +171,7 @@ def update(self) -> None:
         ),
         max_characters=3000,
         max_tokens=2000,
+        terminators=(),
     )
     most_salient_justification = justification_chain_of_thought.open_question(
         question=(
@@ -181,6 +184,7 @@ def update(self) -> None:
         answer_prefix=f'{self._agent_name}',
         max_characters=2000,
         max_tokens=1000,
+        terminators=(),
     )
     salient_justification = (
         f'[thought] {self._agent_name} {most_salient_justification}')

diff --git a/concordia/utils/plotting.py b/concordia/utils/plotting.py
@@ -73,7 +73,7 @@ def plot_df_pie(df: pd.DataFrame,
     group_by: Group data by this field, plot each one in its own figure.
     value: The name of the value to aggregate for the pie chart regions.
   """
-  cmap = mpl.colormaps['Paired']
+  cmap = mpl.colormaps['Paired']  # pylint: disable=unsubscriptable-object
   colours = cmap(range(len(scale)))
   scale_to_colour = dict(zip(scale, colours))
 

diff --git a/examples/whodunnit.ipynb b/examples/whodunnit.ipynb
@@ -329,6 +329,13 @@
         "      ),\n",
         "      name='role playing instructions\\n')\n",
         "\n",
+        "  if agent_config.extras.get('murderer', False):\n",
+        "    fact = generic_components.constant.ConstantComponent(\n",
+        "        state=f'{agent_name} murdered {VICTIM}.', name='fact')\n",
+        "  else:\n",
+        "    fact = generic_components.constant.ConstantComponent(\n",
+        "        state=f'{agent_name} did not kill {VICTIM}.', name='fact')\n",
+        "\n",
         "  time = generic_components.report_function.ReportFunction(\n",
         "      name='Current time',\n",
         "      function=clock.current_time_interval_str,\n",
@@ -434,6 +441,7 @@
         "      agent_config.name,\n",
         "      clock_now=clock.now,\n",
         "      components=[instructions,\n",
+        "                  fact,\n",
         "                  initial_goal_component,\n",
         "                  relevant_memories,\n",
         "                  persona,\n",
@@ -470,6 +478,7 @@
         "      clock=clock,\n",
         "      verbose=True,\n",
         "      components=[instructions,\n",
+        "                  fact,\n",
         "                  persona,\n",
         "                  justification,\n",
         "                  reflection,\n",
@@ -494,7 +503,7 @@
         "      question='What is {opining_player}\\'s opinion of {of_player}?',\n",
         "  )\n",
         "  agent.add_component(reputation_metric)\n",
-        "  return agent\n"
+        "  return agent"
       ]
     },
     {
@@ -533,6 +542,7 @@
         "                 knowledge_of_scandal.get(WHO_FOUND_BODY, '')),\n",
         "        traits = make_random_big_five(),\n",
         "        formative_ages = sorted(random.sample(range(5, 40), 5)),\n",
+        "        extras={'murderer': False},\n",
         "    ),\n",
         "    formative_memories.AgentConfig(\n",
         "        name=MURDERER,\n",
@@ -542,6 +552,7 @@
         "                 knowledge_of_scandal.get(MURDERER, '')),\n",
         "        traits = make_random_big_five(),\n",
         "        formative_ages = sorted(random.sample(range(5, 40), 5)),\n",
+        "        extras={'murderer': True},\n",
         "    ),\n",
         "    formative_memories.AgentConfig(\n",
         "        name='Donald',\n",
@@ -552,6 +563,7 @@
         "                 knowledge_of_scandal.get('Donald', '')),\n",
         "        traits = make_random_big_five(),\n",
         "        formative_ages = sorted(random.sample(range(5, 40), 5)),\n",
+        "        extras={'murderer': False},\n",
         "    ),\n",
         "    formative_memories.AgentConfig(\n",
         "        name='Ellen',\n",
@@ -562,6 +574,7 @@
         "                 knowledge_of_scandal.get('Ellen', '')),\n",
         "        traits = make_random_big_five(),\n",
         "        formative_ages = sorted(random.sample(range(5, 40), 5)),\n",
+        "        extras={'murderer': False},\n",
         "    ),\n",
         "]"
       ]
@@ -682,12 +695,13 @@
       "source": [
         "# @title Create the game master's thought chain\n",
         "account_for_agency_of_others = thought_chains_lib.AccountForAgencyOfOthers(\n",
-        "    model=model, players=players, verbose=True)\n",
-        "\n",
+        "    model=model, players=players, verbose=False)\n",
         "thought_chain = [\n",
+        "    thought_chains_lib.extract_direct_quote,\n",
         "    thought_chains_lib.attempt_to_most_likely_outcome,\n",
         "    thought_chains_lib.result_to_effect_caused_by_active_player,\n",
-        "    account_for_agency_of_others\n",
+        "    account_for_agency_of_others,\n",
+        "    thought_chains_lib.restore_direct_quote,\n",
         "]"
       ]
     },
@@ -828,10 +842,12 @@
         "fig, ax = plt.subplots(1, len(available_channels), figsize=(6, 2))\n",
         "tb = [channel for channel in available_channels]\n",
         "for idx, channel in enumerate(available_channels):\n",
-        "  plotting.plot_line_measurement_channel(measurements, channel,\n",
-        "                                         group_by=group_by[channel],\n",
-        "                                         xaxis='time_str',\n",
-        "                                         ax=ax[idx])\n",
+        "  plotting.plot_line_measurement_channel(\n",
+        "      measurements,\n",
+        "      channel,\n",
+        "      group_by=group_by[channel],\n",
+        "      xaxis='time_str',\n",
+        "      ax=ax[idx])\n",
         "  ax[idx].set_title(channel)\n",
         "\n",
         "fig.set_constrained_layout(constrained=True)"