crfm-helm 0.5.7__py3-none-any.whl → 0.5.8__py3-none-any.whl

helm/benchmark/scenarios/wildbench_scenario.py

@@ -2,11 +2,13 @@ import datasets
 import os
 from typing import List
 
+from helm.benchmark.presentation.taxonomy_info import TaxonomyInfo
 from helm.benchmark.scenarios.scenario import (
     Scenario,
     Instance,
     TEST_SPLIT,
     Input,
+    ScenarioMetadata,
 )
 from helm.common.general import ensure_directory_exists
 
@@ -81,3 +83,19 @@ class WildBenchScenario(Scenario):
             instances.append(instance)
 
         return instances
+
+    def get_metadata(self) -> ScenarioMetadata:
+        return ScenarioMetadata(
+            name=self.name,
+            display_name="WildBench",
+            description=self.description,
+            main_metric="wildbench_score_rescaled",
+            main_split="test",
+            taxonomy=TaxonomyInfo(
+                task="instruction following",
+                what="GPT-judged instruction following with instructions collected from real-user conversations",
+                who="real-world users",
+                when="2024",
+                language="English",
+            ),
+        )

helm/benchmark/scenarios/wmt_14_scenario.py

@@ -1,5 +1,6 @@
 from typing import List, Any
 from datasets import load_dataset
+from helm.benchmark.presentation.taxonomy_info import TaxonomyInfo
 from helm.common.hierarchical_logger import htrack_block
 from helm.benchmark.scenarios.scenario import (
     Scenario,
@@ -11,6 +12,7 @@ from helm.benchmark.scenarios.scenario import (
     CORRECT_TAG,
     Input,
     Output,
+    ScenarioMetadata,
 )
 
 
@@ -106,3 +108,13 @@ class WMT14Scenario(Scenario):
                     )
                 )
         return instances
+
+    def get_metadata(self) -> ScenarioMetadata:
+        return ScenarioMetadata(
+            name="wmt_14",
+            display_name="WMT 2014",
+            description="WMT 2014 is a collection of machine translation datasets.",
+            taxonomy=TaxonomyInfo(task="machine translation", what="n/a", when="n/a", who="n/a", language="English"),
+            main_metric="bleu_4",
+            main_split="test",
+        )

helm/benchmark/static/schema_arabic.yaml

@@ -92,6 +92,12 @@ metrics:
     short_display_name: PEM
     description: Fraction of instances that the predicted output matches the prefix of a correct reference up to light processing.
     lower_is_better: false
+  - name: alrage_score
+    # TODO: should call this prefix_quasi_exact_match
+    display_name: ALRAGE Score
+    short_display_name: Score
+    description: Score of the output judged by GPT-4o.
+    lower_is_better: false
 
 ############################################################
 perturbations: []
@@ -134,17 +140,20 @@ run_groups:
   - name: arabic_scenarios
     display_name: Arabic Scenarios
     description: Arabic Scenarios
-    category: All scenarios
+    category: Scenarios
     subgroups:
-      - mmmlu
-      - arabic_mmlu
       - alghafa
-      - exams_multilingual
+      - arabic_mmlu
+      - arabic_exams
+      - madinah_qa
       - aratrust
+      - alrage
+      - mbzuai_human_translated_arabic_mmlu
 
-  - name: mmmlu
-    display_name: Multilingual MMLU (Arabic)
-    description: Multilingual MMLU (Arabic)
+  - name: mbzuai_human_translated_arabic_mmlu
+    display_name: MBZUAI Human-Translated Arabic MMLU
+    short_display_name: Translated MMLU
+    description: A translation from MBZUAI by human translators of the Massive Multitask Language Understanding benchmark.
     metric_groups:
       - accuracy
       - efficiency
@@ -160,8 +169,8 @@ run_groups:
       language: Arabic
 
   - name: arabic_mmlu
-    display_name: Arabic MMLU
-    description: Arabic MMLU
+    display_name: ArabicMMLU
+    description: ArabicMMLU
     metric_groups:
       - accuracy
       - efficiency
@@ -193,9 +202,9 @@ run_groups:
       when: "before 2023"
       language: Arabic
 
-  - name: exams_multilingual
-    display_name: EXAMS (Arabic)
-    description: EXAMS (Arabic)
+  - name: arabic_exams
+    display_name: Arabic EXAMS
+    description: Arabic EXAMS
     metric_groups:
       - accuracy
       - efficiency
@@ -226,3 +235,37 @@ run_groups:
       who: "academic exams writers and takers"
       when: "before 2024"
       language: Arabic
+
+  - name: alrage
+    display_name: ALRAGE
+    description: ALRAGE
+    metric_groups:
+      - accuracy
+      - efficiency
+      - general_information
+    environment:
+      main_name: alrage_score
+      main_split: test
+    taxonomy:
+      task: "openbook (RAG) open-ended question answering"
+      what: "?"
+      who: "?"
+      when: "?"
+      language: Arabic
+
+  - name: madinah_qa
+    display_name: MadinahQA
+    description: Arabic language competency benchmark
+    metric_groups:
+      - accuracy
+      - efficiency
+      - general_information
+    environment:
+      main_name: exact_match
+      main_split: test
+    taxonomy:
+      task: "question answering"
+      what: "academic questions about Arabic language"
+      who: "academic exams writers and takers"
+      when: "before 2024"
+      language: Arabic

helm/benchmark/static/schema_long_context.yaml

@@ -194,7 +194,7 @@ run_groups:
       - ruler_hotpotqa
       - ruler_squad
       - infinite_bench_en_sum
-      - infinite_bench_en_qa
+      # - infinite_bench_en_qa
       - infinite_bench_en_mc
       - openai_mrcr
 
@@ -233,22 +233,22 @@ run_groups:
       when: Before 2018
       language: English
 
-  - name: infinite_bench_en_qa
-    display_name: ∞Bench En.QA
-    description: ∞Bench En.QA is a open-ended question answering task that requires locating and processing information within a novel, performing reasoning through aggregation or filtering to derive answers. ([Zhang et al., 2024](https://arxiv.org/abs/2402.13718))
-    metric_groups:
-      - accuracy
-      - general_information
-      - annotation_metrics
-    environment:
-      main_name: f1_score
-      main_split: test
-    taxonomy:
-      task: question answering
-      what: Novels
-      who: Novel authors
-      when: Before 2024
-      language: English
+  # - name: infinite_bench_en_qa
+  #   display_name: ∞Bench En.QA
+  #   description: ∞Bench En.QA is a open-ended question answering task that requires locating and processing information within a novel, performing reasoning through aggregation or filtering to derive answers. ([Zhang et al., 2024](https://arxiv.org/abs/2402.13718))
+  #   metric_groups:
+  #     - accuracy
+  #     - general_information
+  #     - annotation_metrics
+  #   environment:
+  #     main_name: f1_score
+  #     main_split: test
+  #   taxonomy:
+  #     task: question answering
+  #     what: Novels
+  #     who: Novel authors
+  #     when: Before 2024
+  #     language: English
 
   - name: infinite_bench_en_mc
     display_name: ∞Bench En.MC

helm/benchmark/static/schema_medhelm.yaml

@@ -484,6 +484,8 @@ run_groups:
       - ehrshot
       - head_qa
       - medbullets
+      - med_qa
+      - med_mcqa
       - medalign
       - shc_ptbm_med
       - shc_sei_med
@@ -657,6 +659,40 @@ run_groups:
       when: Any
       language: English
 
+  - name: med_qa
+    display_name: MedQA
+    description: MedQA is an open domain question answering dataset composed of questions from professional medical board exams ([Jin et al. 2020](https://arxiv.org/pdf/2009.13081.pdf)).
+    metric_groups:
+      - accuracy
+      - efficiency
+      - general_information
+    environment:
+      main_name: exact_match
+      main_split: test
+    taxonomy:
+      task: question answering
+      what: n/a
+      who: n/a
+      when: n/a
+      language: English
+
+  - name: med_mcqa
+    display_name: MedMCQA
+    description: MedMCQA is a "multiple-choice question answering (MCQA) dataset designed to address real-world medical entrance exam questions ([Flores et al. 2020](https://arxiv.org/abs/2203.14371)).
+    metric_groups:
+      - accuracy
+      - efficiency
+      - general_information
+    environment:
+      main_name: exact_match
+      main_split: valid
+    taxonomy:
+      task: question answering
+      what: n/a
+      who: n/a
+      when: n/a
+      language: English
+
   - name: medalign
     display_name: MedAlign
     short_display_name: MedAlign

helm/benchmark/static/schema_slp.yaml

@@ -0,0 +1,219 @@
+############################################################
+metrics:
+  # Infrastructure metrics:
+  - name: num_perplexity_tokens
+    display_name: '# tokens'
+    description: Average number of tokens in the predicted output (for language modeling, the input too).
+  - name: num_bytes
+    display_name: '# bytes'
+    description: Average number of bytes in the predicted output (for language modeling, the input too).
+
+  - name: num_references
+    display_name: '# ref'
+    description: Number of references.
+  - name: num_train_trials
+    display_name: '# trials'
+    description: Number of trials, where in each trial we choose an independent, random set of training instances.
+  - name: estimated_num_tokens_cost
+    display_name: 'cost'
+    description: An estimate of the number of tokens (including prompt and output completions) needed to perform the request.
+  - name: num_prompt_tokens
+    display_name: '# prompt tokens'
+    description: Number of tokens in the prompt.
+  - name: num_prompt_characters
+    display_name: '# prompt chars'
+    description: Number of characters in the prompt.
+  - name: num_completion_tokens
+    display_name: '# completion tokens'
+    description: Actual number of completion tokens (over all completions).
+  - name: num_output_tokens
+    display_name: '# output tokens'
+    description: Actual number of output tokens.
+  - name: max_num_output_tokens
+    display_name: 'Max output tokens'
+    description: Maximum number of output tokens (overestimate since we might stop earlier due to stop sequences).
+  - name: num_requests
+    display_name: '# requests'
+    description: Number of distinct API requests.
+  - name: num_instances
+    display_name: '# eval'
+    description: Number of evaluation instances.
+  - name: num_train_instances
+    display_name: '# train'
+    description: Number of training instances (e.g., in-context examples).
+  - name: prompt_truncated
+    display_name: truncated
+    description: Fraction of instances where the prompt itself was truncated (implies that there were no in-context examples).
+  - name: finish_reason_length
+    display_name: finish b/c length
+    description: Fraction of instances where the the output was terminated because of the max tokens limit.
+  - name: finish_reason_stop
+    display_name: finish b/c stop
+    description: Fraction of instances where the the output was terminated because of the stop sequences.
+  - name: finish_reason_endoftext
+    display_name: finish b/c endoftext
+    description: Fraction of instances where the the output was terminated because the end of text token was generated.
+  - name: finish_reason_unknown
+    display_name: finish b/c unknown
+    description: Fraction of instances where the the output was terminated for unknown reasons.
+  - name: num_completions
+    display_name: '# completions'
+    description: Number of completions.
+  - name: predicted_index
+    display_name: Predicted index
+    description: Integer index of the reference (0, 1, ...) that was predicted by the model (for multiple-choice).
+
+  # Accuracy metrics:
+  - name: exact_match
+    display_name: Exact match
+    short_display_name: EM
+    description: Fraction of instances that the predicted output matches a correct reference exactly.
+    lower_is_better: false
+  - name: classification_macro_f1
+    display_name: Macro-F1
+    description: Population-level macro-averaged F1 score.
+    lower_is_better: false
+  - name: classification_micro_f1
+    display_name: Micro-F1
+    description: Population-level micro-averaged F1 score.
+    lower_is_better: false
+  - name: wer_score
+    display_name: Word Error Rate
+    description: Transcription error rate.
+    lower_is_better: true
+  - name: mer_score
+    display_name: Character Error Rate
+    description: Character error rate.
+    lower_is_better: true
+
+############################################################
+perturbations: []
+
+############################################################
+metric_groups:
+  - name: accuracy
+    display_name: Accuracy
+    hide_win_rates: true
+    metrics:
+      - name: exact_match
+        split: ${main_split}
+      - name: classification_macro_f1
+        split: ${main_split}
+      - name: classification_micro_f1
+        split: ${main_split}
+
+  - name: transcription_accuracy
+    display_name: Transcription Accuracy
+    hide_win_rates: true
+    metrics:
+      - name: wer_score
+        split: ${main_split}
+      - name: mer_score
+        split: ${main_split}
+
+  - name: efficiency
+    display_name: Efficiency
+    metrics:
+    - name: inference_runtime
+      split: ${main_split}
+
+  - name: general_information
+    display_name: General information
+    hide_win_rates: true
+    metrics:
+    - name: num_instances
+      split: ${main_split}
+    - name: num_train_instances
+      split: ${main_split}
+    - name: prompt_truncated
+      split: ${main_split}
+    - name: num_prompt_tokens
+      split: ${main_split}
+    - name: num_output_tokens
+      split: ${main_split}
+
+############################################################
+
+run_groups:
+  - name: slp
+    display_name: SLP Scenarios
+    description: SLP-language scenarios
+    category: All scenarios
+    subgroups:
+      - disorder_diagnosis
+      - transcription
+      - symptom_diagnosis
+      - disorder_type_diagnosis
+
+
+  - name: disorder_diagnosis
+    display_name: Disorder Diagnosis Accuracy
+    description: >
+      Macro-averaged accuracy on disorder diagnosis for pediatric speech disorder.
+    metric_groups:
+      - accuracy
+      - efficiency
+      - general_information
+    environment:
+      main_name: classification_micro_f1
+      main_split: test
+    taxonomy:
+      task: classification
+      what: n/a
+      who: n/a
+      when: "?"
+      language: English
+
+  - name: transcription
+    display_name: Transcription Accuracy
+    description: >
+      Model transcription accuracy on understanding disordered pediatric speech
+    metric_groups:
+      - transcription_accuracy
+      - efficiency
+      - general_information
+    environment:
+      main_name: wer_score
+      main_split: test
+    taxonomy:
+      task: transcription
+      what: disordered pediatric speech
+      who: n/a
+      when: "?"
+      language: English
+
+  - name: symptom_diagnosis
+    display_name: Symptom Diagnosis Accuracy
+    description: >
+      Macro-averaged accuracy on symptom diagnosis for pediatric speech disorder.
+    metric_groups:
+      - accuracy
+      - efficiency
+      - general_information
+    environment:
+      main_name: classification_micro_f1
+      main_split: test
+    taxonomy:
+      task: classification
+      what: n/a
+      who: n/a
+      when: "?"
+      language: English
+
+  - name: disorder_type_diagnosis
+    display_name: Disorder Type Diagnosis Accuracy
+    description: >
+      Macro-averaged accuracy on disorder type diagnosis for pediatric speech disorder.
+    metric_groups:
+      - accuracy
+      - efficiency
+      - general_information
+    environment:
+      main_name: classification_micro_f1
+      main_split: test
+    taxonomy:
+      task: classification
+      what: n/a
+      who: n/a
+      when: "?"
+      language: English