crfm-helm 0.5.1__py3-none-any.whl → 0.5.3__py3-none-any.whl

helm/benchmark/static/schema_call_center.yaml

@@ -0,0 +1,232 @@
+---
+############################################################
+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
+
+  # Summarization metrics:
+  - name: annotation_call_center_summarization_faithfulness
+    display_name: Faithfulness
+    short_display_name: Faithfulness
+    description: Whether all the information expressed by the summary can be inferred from the source transcript.
+    lower_is_better: false
+  - name: annotation_call_center_summarization_relevance
+    display_name: Relevance
+    short_display_name: Relevance
+    description: Whether the summary includes only important information from the source.
+    lower_is_better: false
+  - name: annotation_call_center_summarization_coherence
+    display_name: Coherence
+    short_display_name: Coherence
+    description: Whether the summary organizes the relevant information into a well-structured summary.
+    lower_is_better: false
+
+  - name: annotation_call_center_summarization_pairwise_comparison_score
+    display_name: Pairwise
+    short_display_name: Pairwise
+    description: Whether the model's summary was preferred by the evaluator model
+    lower_is_better: false
+    
+  - name: annotation_call_center_summarization_key_points_recall_score
+    display_name: Recall
+    short_display_name: Recall
+    description: How many key items were recalled
+    lower_is_better: false
+############################################################
+perturbations: []
+
+############################################################
+metric_groups:
+  - name: summarization_metrics
+    display_name: Summarization
+    hide_win_rates: true
+    metrics:
+      - name: annotation_call_center_summarization_faithfulness
+        split: ${main_split}
+      - name: annotation_call_center_summarization_relevance
+        split: ${main_split}
+      - name: annotation_call_center_summarization_coherence
+        split: ${main_split}
+
+  - name: pairwise_comparison_metrics
+    display_name: Pairwise Comparison
+    hide_win_rates: true
+    metrics:
+      - name: annotation_call_center_summarization_pairwise_comparison_score
+        split: ${main_split}
+
+  - name: key_points_recall_metrics
+    display: Key Points Recall
+    hide_win_rates: true
+    metrics:
+      - name: annotation_call_center_summarization_key_points_recall_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: call_center_scenarios
+    display_name: Call Center Scenarios
+    description: Scenarios representating realistic tasks from the call center.
+    category: All scenarios
+    subgroups:
+      - call_center_summarization
+      - call_center_summarization_real_call_transcripts
+      - call_center_summarization_pairwise_comparison
+      - call_center_summarization_key_points_recall
+
+  - name: call_center_summarization
+    display_name: Summarization
+    description: summarization
+    metric_groups:
+      # - accuracy
+      - summarization_metrics
+      - efficiency
+      - general_information
+    environment:
+      main_split: test
+    taxonomy:
+      task: summarization
+      what: n/a
+      who: n/a
+      when: "?"
+      language: English
+
+  - name: call_center_summarization_real_call_transcripts
+    display_name: Summarization (Real)
+    description: Summarization with real call transcripts
+    metric_groups:
+      # - accuracy
+      - summarization_metrics
+      - efficiency
+      - general_information
+    environment:
+      main_split: test
+    taxonomy:
+      task: summarization
+      what: n/a
+      who: n/a
+      when: "?"
+      language: English
+
+  - name: call_center_summarization_pairwise_comparison
+    display_name: Summarization (Pairwise)
+    description: summarization
+    metric_groups:
+      # - accuracy
+      - pairwise_comparison_metrics
+      - efficiency
+      - general_information
+    environment:
+      main_split: test
+    taxonomy:
+      task: summarization
+      what: n/a
+      who: n/a
+      when: "?"
+      language: English
+
+  - name: call_center_summarization_key_points_recall
+    display_name: Summarization (Key Points Recall)
+    description: summarization
+    metric_groups:
+      # - accuracy
+      - key_points_recall_metrics
+      - efficiency
+      - general_information
+    environment:
+      main_split: test
+    taxonomy:
+      task: summarization
+      what: n/a
+      who: n/a
+      when: "?"
+      language: English

helm/benchmark/static/schema_classic.yaml

@@ -1,64 +1,8 @@
 ---
 ############################################################
-adapter:
-  - name: method
-    description: The high-level strategy for converting instances into a prompt for the language model.
-    values:
-      - name: generation
-        description: Given the input, the model generates the output free-form.
-      - name: multiple_choice_joint
-        description: Given the input, the model selects from multiple-choice options (A., B., C., D., E.).
-      - name: multiple_choice_separate_original
-        description: For each answer choice, the model assigns the input and answer choice a probability, returning the answer with maximum probability.
-      - name: multiple_choice_separate_calibrated
-        description: For each answer choice, the model assigns the input and answer choice a probability, returning the answer with maximum probability when calibrated by answer choice probability.
-      - name: language_modeling
-        description: Given the input, the model assigns the sequence a probability.
-  - name: instructions
-    description: The description of the task that is included at the very beginning of the prompt.
-  - name: global_prefix
-    description: The string that is prepended to the prompt.
-  - name: instance_prefix
-    description: The string that is included before each instance (e.g., '\n\n').
-  - name: input_prefix
-    description: The string that is included before each input (e.g., 'Question:').
-  - name: input_suffix
-    description: The string that is included after each input (e.g., '\n').
-  - name: reference_prefix
-    description: The string that is included before each reference (for multiple-choice questions).
-  - name: reference_suffix
-    description: The string that is included after each reference (for multiple-choice questions).
-  - name: output_prefix
-    description: The string that is included before the correct answer/predicted output (e.g., 'Answer:').
-  - name: output_suffix
-    description: The string that is included after the correct answer/predicted output (e.g., '\n').
-  - name: substitutions
-    description: A list of regular expression substitutions (e.g., replacing '\n' with ';\n') to perform at the very end on the prompt.
-  - name: max_train_instances
-    description: Maximum number of training instances to include in the prompt (currently by randomly sampling).
-  - name: max_eval_instances
-    description: Maximum number of instances to evaluate on (over all splits - test, valid, etc.).
-  - name: num_outputs
-    description: Maximum number of possible outputs to generate by sampling multiple outputs.
-  - name: num_train_trials
-    description: Number of trials, where in each trial we choose an independent, random set of training instances. Used to compute variance.
-  - name: sample_train
-    description: If true, randomly sample N training examples; if false, select N consecutive training examples
-  - name: model
-    description: Name of the language model (<creator_organization>/<model name>) to send requests to.
-  - name: model_deployment
-    description: Name of the language model deployment (<host_organization>/<model name>) to send requests to.
-  - name: temperature
-    description: Temperature parameter used in generation.
-  - name: max_tokens
-    description: Maximum number of tokens to generate.
-  - name: stop_sequences
-    description: List of sequences, where we stop generation if we encounter any of them.
-  - name: random
-    description: Random seed (string), which guarantees reproducibility.
-  - name: multi_label
-    description: If true, for instances with multiple correct reference, the gold answer should be considered to be all of the correct references rather than any of the correct references.
-
+# For backwards compatibility with older versions of HELM.
+# TODO: Remove this after 2024-09-01.
+adapter: []
 ############################################################
 metrics:
   # Infrastructure metrics: