crfm-helm 0.5.3__py3-none-any.whl → 0.5.5__py3-none-any.whl

helm/benchmark/static/schema_czech_bank.yaml

@@ -0,0 +1,148 @@
+---
+############################################################
+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: program_accuracy
+    display_name: Program Accuracy
+    description: Accuracy of the generated programs
+    lower_is_better: false
+  - name: execution_accuracy
+    display_name: Execution Accuracy
+    description: Accuracy of the final result of the generated program
+    lower_is_better: false
+  - name: annotation_financebench_label_correct_answer
+    display_name: Correct Answer
+    description: Whether the final result was correct, as judged by a GPT-4o
+    lower_is_better: false
+  - name: quasi_exact_match
+    display_name: Quasi-exact match
+    short_display_name: EM
+    description: Fraction of instances that the predicted output matches a correct reference up to light processing.
+    lower_is_better: false
+  - name: error_rate
+    display_name: SQL Error Rate
+    short_display_name: SQL Error Rate
+    description: Fraction of generated queries that result in a SQL execution error
+    lower_is_better: true
+
+############################################################
+perturbations: []
+
+############################################################
+metric_groups:
+  - name: accuracy
+    display_name: Accuracy
+    hide_win_rates: true
+    metrics:
+      - name: ${main_name}
+        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: financial_scenarios
+    display_name: Financial Scenarios
+    description: Scenarios for the financial domain
+    category: All scenarios
+    subgroups:
+      - czech_bank_qa
+
+  - name: czech_bank_qa
+    display_name: CzechBankQA
+    description: The CzechBankQA
+    metric_groups:
+      - accuracy
+      - efficiency
+      - general_information
+    environment:
+      main_name: error_rate
+      main_split: test
+    taxonomy:
+      task: text-to-SQL
+      what: queries from financial experts
+      who: financial experts
+      when: "1999"
+      language: English

helm/benchmark/static/schema_enem_challenge.yaml

@@ -0,0 +1,146 @@
+############################################################
+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: quasi_exact_match
+    display_name: Quasi-exact match
+    short_display_name: EM
+    description: Fraction of instances that the predicted output matches a correct reference up to light processing.
+    lower_is_better: false
+  - name: prefix_exact_match
+    display_name: Prefix exact match
+    short_display_name: PEM
+    description: Fraction of instances that the predicted output matches the prefix of a correct reference exactly.
+    lower_is_better: false
+  - name: quasi_prefix_exact_match
+    # TODO: should call this prefix_quasi_exact_match
+    display_name: Prefix quasi-exact match
+    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
+
+
+############################################################
+perturbations: []
+
+############################################################
+metric_groups:
+  - name: accuracy
+    display_name: Accuracy
+    metrics:
+      - name: ${main_name}
+        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: core_scenarios
+    display_name: Core Scenarios
+    description: Core Scenarios
+    category: All scenarios
+    subgroups:
+      - enem_challenge
+
+  - name: enem_challenge
+    display_name: ENEM Challenge
+    description: ENEM Challenge
+    metric_groups:
+      - accuracy
+    # - efficiency
+      - general_information
+    environment:
+      main_name: exact_match
+      main_split: test
+    taxonomy:
+      task: "multiple-choice question answering"
+      what: "general academic subjects"
+      who: "brazilian ministry of education"
+      when: "between 2009 and 2023"
+      language: Portuguese

helm/benchmark/static/schema_enterprise.yaml

@@ -0,0 +1,298 @@
+---
+############################################################
+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: num_prompt_tokens
+    display_name: '# prompt tokens'
+    description: Number of tokens 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: 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.
+  # 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: quasi_exact_match
+    display_name: Quasi-exact match
+    short_display_name: EM
+    description: Fraction of instances that the predicted output matches a correct reference up to light processing.
+    lower_is_better: false
+  - name: rouge_1
+    display_name: ROUGE-1
+    description: Average ROUGE score [(Lin, 2004)](https://aclanthology.org/W04-1013/) based on 1-gram overlap.
+    lower_is_better: false
+  - name: rouge_2
+    display_name: ROUGE-2
+    description: Average ROUGE score [(Lin, 2004)](https://aclanthology.org/W04-1013/) based on 2-gram overlap.
+    lower_is_better: false
+  - name: rouge_l
+    display_name: ROUGE-L
+    description: Average ROUGE score [(Lin, 2004)](https://aclanthology.org/W04-1013/) based on longest common subsequence overlap.
+    lower_is_better: false
+  - name: classification_weighted_f1
+    display_name: Weighted F1
+    description: Weighted F1 score
+    lower_is_better: false
+  - name: float_equiv
+    display_name: Float Equivalence
+    description: Float Equivalence
+    lower_is_better: false
+
+############################################################
+perturbations: []
+
+############################################################
+metric_groups:
+  - name: accuracy
+    display_name: Accuracy
+    metrics:
+      - name: ${main_name}
+        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: financial_scenarios
+    display_name: Financial Scenarios
+    description: Scenarios for the financial domain
+    category: All scenarios
+    subgroups:
+      - gold_commodity_news
+      - financial_phrasebank
+      - conv_fin_qa_calc
+
+  - name: legal_scenarios
+    display_name: Legal Scenarios
+    description: Scenarios for the legal domain
+    category: All scenarios
+    subgroups:
+      - legal_contract_summarization
+      - casehold
+      - echr_judgment_classification
+      - legal_opinion_sentiment_classification
+
+  - name: climate_scenarios
+    display_name: Climate Scenarios
+    description: Scenarios for the climate domain
+    category: All scenarios
+    subgroups:
+      - sumosum
+
+  - name: cyber_security_scenarios
+    display_name: Cyber Security Scenarios
+    description: Scenarios for the cyber security domain
+    category: All scenarios
+    subgroups:
+      - cti_to_mitre
+
+  - name: financial_phrasebank
+    display_name: Financial Phrasebank (Sentiment Classification)
+    description: A sentiment classification benchmark based on the dataset from Good Debt or Bad Debt - Detecting Semantic Orientations in Economic Texts [(Malo et al., 2013)](https://arxiv.org/abs/1307.5336).
+    metric_groups:
+      - accuracy
+      - efficiency
+      - general_information
+    environment:
+      main_name: classification_weighted_f1
+      main_split: test
+    taxonomy:
+      task: sentiment analysis
+      what: phrases from financial news texts and company press releases
+      who: annotators with adequate business education background
+      when: before 2013
+      language: English
+
+  - name: conv_fin_qa_calc
+    display_name: ConvFinQACalc
+    description: "A mathematical calculation benchmark based on ConvFinQA: Exploring the Chain of Numerical Reasoning in Conversational Finance Question Answering [(Chen ey al., 2022)](https://arxiv.org/pdf/2210.03849.pdf)."
+    metric_groups:
+      - accuracy
+      - efficiency
+      - general_information
+    environment:
+      main_name: float_equiv
+      main_split: valid
+    taxonomy:
+      task: question answering with numeric reasoning
+      what: financial reports
+      who: financial experts
+      when: 1999 to 2019
+      language: English
+
+  - name: gold_commodity_news
+    display_name: Gold Commodity News
+    description: A classification benchmark based on a dataset of human-annotated gold commodity news headlines ([Sinha & Khandait, 2019](https://arxiv.org/abs/2009.04202)).
+    metric_groups:
+      - accuracy
+      - efficiency
+      - general_information
+    environment:
+      main_name: classification_weighted_f1
+      main_split: test
+    taxonomy:
+      task: text classification
+      what: gold commodity news headlines
+      who: financial journalists
+      when: 2000-2019
+      language: English
+
+  - name: legal_contract_summarization
+    display_name: Legal Contract Summarization
+    description: Plain English Summarization of Contracts [(Manor et al., 2019)](https://aclanthology.org/W19-2201.pdf).
+    metric_groups:
+      - accuracy
+      - efficiency
+      - general_information
+    environment:
+      main_name: rouge_l
+      main_split: test
+    taxonomy:
+      task: summarization
+      what: legal contracts (e.g. terms of service, license agreements)
+      who: lawyers
+      when: before 2019
+      language: English
+
+  - name: casehold
+    display_name: CaseHOLD
+    description: CaseHOLD (Case Holdings On Legal Decisions) is a multiple choice question answering scenario where the task is to identify the relevant holding of a cited case [(Zheng et al, 2021)](https://arxiv.org/pdf/2104.08671.pdf).
+    metric_groups:
+      - accuracy
+      - efficiency
+      - general_information
+    environment:
+      main_name: quasi_exact_match
+      main_split: test
+    taxonomy:
+      task: question answering
+      what: Harvard Law Library case law corpus
+      who: legal professionals
+      when: before 2021
+      language: English
+
+  - name: echr_judgment_classification
+    display_name: ECHR Judgment Classification
+    description: The "Binary Violation" Classification task from the paper Neural Legal Judgment Prediction in English [(Chalkidis et al., 2019)](https://arxiv.org/pdf/1906.02059.pdf). The task is to analyze the description of a legal case from the European Court of Human Rights (ECHR), and classify it as positive if any human rights article or protocol has been violated and negative otherwise.
+    metric_groups:
+      - accuracy
+      - general_information
+    environment:
+      main_name: classification_weighted_f1
+      main_split: test
+    taxonomy:
+      task: text classification
+      what: casees from the European Court of Human Rights
+      who: judiciary of the European Court of Human Rights
+      when: 2014-2018 (train) and 2014-2018 (test)
+      language: English
+
+  - name: legal_opinion_sentiment_classification
+    display_name: Legal Opinion Sentiment Classification
+    description: A legal opinion sentiment classification task based on the paper Effective Approach to Develop a Sentiment Annotator For Legal Domain in a Low Resource Setting [(Ratnayaka et al., 2020)](https://arxiv.org/pdf/2011.00318.pdf).
+    metric_groups:
+      - accuracy
+      - general_information
+    environment:
+      main_name: quasi_exact_match
+      main_split: test
+    taxonomy:
+      task: sentiment analysis
+      what: United States legal opinion texts
+      who: United States courts
+      when: Before 2020
+      language: English
+
+  - name: sumosum
+    display_name: SUMO Web Claims Summarization
+    description: A summarization benchmark based on the climate subset of the SUMO dataset ([Mishra et al., 2020](https://aclanthology.org/2020.wnut-1.12/)).
+    metric_groups:
+      - accuracy
+      - efficiency
+      - general_information
+    environment:
+      main_name: rouge_l
+      main_split: test
+    taxonomy:
+      task: summarization
+      what: Articles from climatefeedback.org
+      who: Writers of news articles and web documents
+      when: Before 2020
+      language: English
+      main_name: quasi_exact_match
+      main_split: test
+
+  - name: cti_to_mitre
+    display_name: CTI-to-MITRE Cyber Threat Intelligence
+    description: A classification benchmark based on Automatic Mapping of Unstructured Cyber Threat Intelligence - An Experimental Study [(Orbinato et al., 2022)](https://arxiv.org/pdf/2208.12144.pdf).
+    metric_groups:
+      - accuracy
+      - efficiency
+      - general_information
+    environment:
+      main_name: quasi_exact_match
+      main_split: test
+    taxonomy:
+      task: text classification
+      what: Descriptions of malicious techniques
+      who: Security professionals
+      when: Before 2022
+      language: English

helm/benchmark/static/schema_finance.yaml

@@ -83,6 +83,14 @@ metrics:
     description: Fraction of instances that the predicted output matches a correct reference up to light processing.
     lower_is_better: false
 
+  # Efficiency metrics:
+  - name: inference_runtime
+    display_name: Observed inference runtime (s)
+    short_display_name: Observed inference time (s)
+    lower_is_better: true
+    description: Average observed time to process a request to the model (via an API, and thus depends on particular deployment).
+
+
 ############################################################
 perturbations: []
 
@@ -90,12 +98,16 @@ perturbations: []
 metric_groups:
   - name: accuracy
     display_name: Accuracy
+    aggregation_strategies: 
+      - mean
     metrics:
       - name: ${main_name}
         split: ${main_split}
 
   - name: efficiency
     display_name: Efficiency
+    aggregation_strategies: 
+      - mean
     metrics:
     - name: inference_runtime
       split: ${main_split}
@@ -145,7 +157,7 @@ run_groups:
 
   - name: financebench
     display_name: FinanceBench
-    description: FinanceBench is a benchmark for open book financial question answering. It comprises 10,231 questions about publicly traded companies, with corresponding answers and evidence strings
+    description: FinanceBench is a benchmark for open book financial question answering. It comprises 10,231 questions about publicly traded companies, with corresponding answers and evidence strings [(Islam et al., 2023)](https://arxiv.org/abs/2311.11944/).
     metric_groups:
       - accuracy
       - efficiency
@@ -163,7 +175,7 @@ run_groups:
   - name: banking77
     display_name: BANKING77
     short_display_name: BANKING77
-    description: BANKING77 is a benchmark for intent classification of customer service queries in the banking domain [(Casanueva et al., 2020)](https://aclanthology.org/2020.nlp4convai-1.5/)).
+    description: BANKING77 is a benchmark for intent classification of customer service queries in the banking domain [(Casanueva et al., 2020)](https://aclanthology.org/2020.nlp4convai-1.5/).
     metric_groups:
       - accuracy
       - efficiency
@@ -177,13 +189,3 @@ run_groups:
       who: banking customers
       when: During or before 2020
       language: English
-
-  # - name: financial_scenarios_ablations
-  #   display_name: Financial Scenarios Ablations
-  #   description: Scenarios for the financial domain with ablations
-  #   category: All scenarios
-  #   subgroups:
-  #     - fin_qa
-  #   adapter_keys_shown:
-  #     - model
-  #     - max_train_instances