Save intermediate outputs from evals, and display it

app/web_ui/src/lib/api_schema.d.ts

@@ -995,11 +995,11 @@ export interface components {
             /** Mean Normalized Squared Error */
             mean_normalized_squared_error: number;
             /** Spearman Correlation */
-            spearman_correlation: number;
+            spearman_correlation: number | null;
             /** Pearson Correlation */
-            pearson_correlation: number;
+            pearson_correlation: number | null;
             /** Kendalltau Correlation */
-            kendalltau_correlation: number;
+            kendalltau_correlation: number | null;
         };
         /**
          * CreateDatasetSplitRequest
@@ -1511,6 +1511,13 @@ export interface components {
              * @description The output of the task. JSON formatted for structured output, plaintext for unstructured output.
              */
             output: string;
+            /**
+             * Intermediate Outputs
+             * @description The intermediate outputs of the task.
+             */
+            intermediate_outputs?: {
+                [key: string]: string;
+            } | null;
             /**
              * Scores
              * @description The scores of the evaluator (specifically the EvalConfig this object is a child of).
@@ -1813,7 +1820,7 @@ export interface components {
          *     Where models have instruct and raw versions, instruct is default and raw is specified.
          * @enum {string}
          */
-        ModelName: "llama_3_1_8b" | "llama_3_1_70b" | "llama_3_1_405b" | "llama_3_2_1b" | "llama_3_2_3b" | "llama_3_2_11b" | "llama_3_2_90b" | "llama_3_3_70b" | "gpt_4o_mini" | "gpt_4o" | "phi_3_5" | "phi_4" | "mistral_large" | "mistral_nemo" | "gemma_2_2b" | "gemma_2_9b" | "gemma_2_27b" | "claude_3_5_haiku" | "claude_3_5_sonnet" | "gemini_1_5_flash" | "gemini_1_5_flash_8b" | "gemini_1_5_pro" | "gemini_2_0_flash" | "nemotron_70b" | "mixtral_8x7b" | "qwen_2p5_7b" | "qwen_2p5_72b" | "deepseek_3" | "deepseek_r1" | "mistral_small_3" | "deepseek_r1_distill_qwen_32b" | "deepseek_r1_distill_llama_70b" | "deepseek_r1_distill_qwen_14b" | "deepseek_r1_distill_qwen_1p5b" | "deepseek_r1_distill_qwen_7b" | "deepseek_r1_distill_llama_8b";
+        ModelName: "llama_3_1_8b" | "llama_3_1_70b" | "llama_3_1_405b" | "llama_3_2_1b" | "llama_3_2_3b" | "llama_3_2_11b" | "llama_3_2_90b" | "llama_3_3_70b" | "gpt_4o_mini" | "gpt_4o" | "phi_3_5" | "phi_4" | "mistral_large" | "mistral_nemo" | "gemma_2_2b" | "gemma_2_9b" | "gemma_2_27b" | "claude_3_5_haiku" | "claude_3_5_sonnet" | "gemini_1_5_flash" | "gemini_1_5_flash_8b" | "gemini_1_5_pro" | "gemini_2_0_flash" | "nemotron_70b" | "mixtral_8x7b" | "qwen_2p5_7b" | "qwen_2p5_72b" | "deepseek_3" | "deepseek_r1" | "mistral_small_3" | "deepseek_r1_distill_qwen_32b" | "deepseek_r1_distill_llama_70b" | "deepseek_r1_distill_qwen_14b" | "deepseek_r1_distill_qwen_1p5b" | "deepseek_r1_distill_qwen_7b" | "deepseek_r1_distill_llama_8b" | "dolphin_2_9_8x22b";
         /**
          * ModelProviderName
          * @description Enumeration of supported AI model providers.

app/web_ui/src/lib/types.ts

@@ -30,3 +30,4 @@ export type EvalResultSummary = components["schemas"]["EvalResultSummary"]
 export type EvalRunResult = components["schemas"]["EvalRunResult"]
 export type EvalConfigCompareSummary =
   components["schemas"]["EvalConfigCompareSummary"]
+export type EvalRun = components["schemas"]["EvalRun"]

app/web_ui/src/routes/(app)/evals/[project_id]/[task_id]/[eval_id]/[eval_config_id]/[run_config_id]/run_result/+page.svelte

@@ -6,6 +6,7 @@
     EvalRunResult,
     Eval,
     EvalConfig,
+    EvalRun,
     TaskRunConfig,
   } from "$lib/types"
   import { client } from "$lib/api_client"
@@ -29,6 +30,8 @@
   let results_error: KilnError | null = null
   let results_loading = true
   let peek_dialog: Dialog | null = null
+  let thinking_dialog: Dialog | null = null
+  let displayed_result: EvalRun | null = null
 
   onMount(async () => {
     peek_dialog?.show()
@@ -179,8 +182,8 @@
       <table class="table">
         <thead>
           <tr>
-            <th>Input</th>
-            <th>Output</th>
+            <th>Input & Output</th>
+            <th>Thinking</th>
             {#each results.eval.output_scores as score}
               <th class="text-center">
                 {score.name}
@@ -192,8 +195,47 @@
         <tbody>
           {#each results.results as result}
             <tr>
-              <td> {result.input} </td>
-              <td> {result.output} </td>
+              <td>
+                <div class="font-medium">Input:</div>
+                <div>
+                  {result.input}
+                </div>
+                <div class="font-medium mt-4">Output:</div>
+                <div>
+                  {result.output}
+                </div>
+              </td>
+              <td>
+                {#if result.intermediate_outputs?.reasoning || result.intermediate_outputs?.chain_of_thought}
+                  <div class="max-w-[600px] min-w-[200px]">
+                    <div class="max-h-[140px] overflow-y-hidden relative">
+                      {result.intermediate_outputs?.reasoning ||
+                        result.intermediate_outputs?.chain_of_thought ||
+                        "N/A"}
+                      <div class="absolute bottom-0 left-0 w-full">
+                        <div
+                          class="h-36 bg-gradient-to-t from-white to-transparent"
+                        ></div>
+                        <div
+                          class="text-center bg-white font-medium font-sm text-gray-500"
+                        >
+                          <button
+                            class="text-gray-500"
+                            on:click={() => {
+                              displayed_result = result
+                              thinking_dialog?.show()
+                            }}
+                          >
+                            See all
+                          </button>
+                        </div>
+                      </div>
+                    </div>
+                  </div>
+                {:else}
+                  N/A
+                {/if}
+              </td>
               {#each results.eval.output_scores as score}
                 {@const score_value =
                   result.scores[string_to_json_key(score.name)]}
@@ -245,3 +287,20 @@
     </div>
   </div>
 </Dialog>
+
+<Dialog
+  bind:this={thinking_dialog}
+  title="Thinking Output"
+  action_buttons={[
+    {
+      label: "Close",
+      isCancel: true,
+    },
+  ]}
+>
+  <div class="font-light text-sm whitespace-pre-wrap">
+    {displayed_result?.intermediate_outputs?.reasoning ||
+      displayed_result?.intermediate_outputs?.chain_of_thought ||
+      "N/A"}
+  </div>
+</Dialog>

libs/core/kiln_ai/adapters/eval/base_eval.py

@@ -1,5 +1,6 @@
 import json
 from abc import abstractmethod
+from typing import Dict
 
 from kiln_ai.adapters.adapter_registry import adapter_for_task
 from kiln_ai.adapters.ml_model_list import ModelProviderName
@@ -40,7 +41,9 @@ def model_and_provider(self) -> tuple[str, ModelProviderName]:
 
         return model_name, ModelProviderName(provider)
 
-    async def run_task_and_eval(self, input: str) -> tuple[TaskRun, EvalScores]:
+    async def run_task_and_eval(
+        self, input: str
+    ) -> tuple[TaskRun, EvalScores, Dict[str, str] | None]:
         if self.run_config is None:
             raise ValueError("Run config is required for run_task_and_eval")
 
@@ -59,14 +62,17 @@ async def run_task_and_eval(self, input: str) -> tuple[TaskRun, EvalScores]:
         # we don't save by default here. We'll save manually after validating the output
         run_output = await run_adapter.invoke(parsed_input)
 
-        eval_output = await self.run_eval(run_output)
+        eval_output, intermediate_outputs = await self.run_eval(run_output)
         validate_schema(eval_output, self.score_schema)
 
-        return run_output, eval_output
+        return run_output, eval_output, intermediate_outputs
 
     @abstractmethod
-    # Runs the eval on the given task run and returns a dictionary of scores which should conform to the score schema
-    async def run_eval(self, task_run: TaskRun) -> EvalScores:
+    # Runs the eval on the given task run
+    # Returns a dictionary of scores which should conform to the score schema, and a dictionary of intermediate outputs
+    async def run_eval(
+        self, task_run: TaskRun
+    ) -> tuple[EvalScores, Dict[str, str] | None]:
         pass
 
     @classmethod

libs/core/kiln_ai/adapters/eval/eval_runner.py

@@ -1,4 +1,5 @@
 import asyncio
+import logging
 from dataclasses import dataclass
 from typing import AsyncGenerator, Dict, List, Literal, Set
 
@@ -10,6 +11,8 @@
 from kiln_ai.datamodel.task import TaskRunConfig
 from kiln_ai.datamodel.task_run import TaskRun
 
+logger = logging.getLogger(__name__)
+
 
 @dataclass
 class EvalJob:
@@ -227,15 +230,18 @@ async def run_job(self, job: EvalJob) -> bool:
 
             task_output: str | None = None
             scores: EvalScores | None = None
+            intermediate_outputs: Dict[str, str] | None = None
             if job.type == "eval_config_eval":
                 # Eval config eval, we use the saved input from the task run, not invoking the task again
-                scores = await evaluator.run_eval(job.item)
+                scores, intermediate_outputs = await evaluator.run_eval(job.item)
                 task_output = job.item.output.output
             else:
                 # Task run eval, we invoke the task again to get a fresh output
-                result_task_run, scores = await evaluator.run_task_and_eval(
-                    job.item.input
-                )
+                (
+                    result_task_run,
+                    scores,
+                    intermediate_outputs,
+                ) = await evaluator.run_task_and_eval(job.item.input)
                 task_output = result_task_run.output.output
 
             # Save the job result
@@ -249,10 +255,11 @@ async def run_job(self, job: EvalJob) -> bool:
                 scores=scores,
                 input=job.item.input,
                 output=task_output,
+                intermediate_outputs=intermediate_outputs,
             )
             eval_run.save_to_file()
 
             return True
         except Exception as e:
-            print(f"Error running eval job for dataset item {job.item.id}: {e}")
+            logger.error(f"Error running eval job for dataset item {job.item.id}: {e}")
             return False

libs/core/kiln_ai/adapters/eval/g_eval.py

@@ -88,7 +88,9 @@ def __init__(self, eval_config: EvalConfig, run_config: RunConfig | None):
 
         self.geval_task = GEvalTask(eval_config)
 
-    async def run_eval(self, task_run: TaskRun) -> EvalScores:
+    async def run_eval(
+        self, task_run: TaskRun
+    ) -> tuple[EvalScores, Dict[str, str] | None]:
         """
         Run this G-Eval on the given task run.
         """
@@ -128,9 +130,11 @@ async def run_eval(self, task_run: TaskRun) -> EvalScores:
         _, run_output = await adapter.invoke_returning_run_output(input)
 
         if self.eval_config.config_type == EvalConfigType.llm_as_judge:
-            return self.build_llm_as_judge_score(run_output)
+            return self.build_llm_as_judge_score(
+                run_output
+            ), run_output.intermediate_outputs
         else:
-            return self.build_g_eval_score(run_output)
+            return self.build_g_eval_score(run_output), run_output.intermediate_outputs
 
     def build_llm_as_judge_score(self, run_output: RunOutput) -> EvalScores:
         """

libs/core/kiln_ai/adapters/eval/test_eval_runner.py

@@ -1,3 +1,4 @@
+from typing import Dict
 from unittest.mock import AsyncMock, patch
 
 import pytest
@@ -484,12 +485,17 @@ async def test_run_job_success_task_run_eval(
         input="test input",
         input_source=data_source,
         output=TaskOutput(output="evaluated output"),
+        intermediate_outputs={"intermediate_output": "intermediate output"},
     )
     mock_scores = {"accuracy": 0.95}
 
     class MockEvaluator(BaseEval):
         async def run_task_and_eval(self, input_text):
-            return mock_result_run, mock_scores
+            return (
+                mock_result_run,
+                mock_scores,
+                {"intermediate_output": "intermediate output"},
+            )
 
     with patch(
         "kiln_ai.adapters.eval.eval_runner.eval_adapter_from_type",
@@ -508,6 +514,9 @@ async def run_task_and_eval(self, input_text):
     assert saved_run.scores == mock_scores
     assert saved_run.input == "test input"
     assert saved_run.output == "evaluated output"
+    assert saved_run.intermediate_outputs == {
+        "intermediate_output": "intermediate output"
+    }
     assert saved_run.parent_eval_config().id == mock_eval_config.id
     assert saved_run.eval_config_eval is False
 
@@ -544,8 +553,10 @@ class MockEvaluator(BaseEval):
         async def run_task_and_eval(self, input_text):
             raise ValueError("Attempted to run task and eval for a config eval")
 
-        async def run_eval(self, task_run: TaskRun) -> EvalScores:
-            return mock_scores
+        async def run_eval(
+            self, task_run: TaskRun
+        ) -> tuple[EvalScores, Dict[str, str] | None]:
+            return mock_scores, {"intermediate_output": "intermediate output"}
 
     with patch(
         "kiln_ai.adapters.eval.eval_runner.eval_adapter_from_type",

libs/core/kiln_ai/datamodel/eval.py

@@ -101,6 +101,10 @@ class EvalRun(KilnParentedModel):
     output: str = Field(
         description="The output of the task. JSON formatted for structured output, plaintext for unstructured output."
     )
+    intermediate_outputs: Dict[str, str] | None = Field(
+        default=None,
+        description="The intermediate outputs of the task.",
+    )
     scores: EvalScores = Field(
         description="The scores of the evaluator (specifically the EvalConfig this object is a child of)."
     )