crfm-helm 0.4.0__py3-none-any.whl → 0.5.0__py3-none-any.whl

helm/benchmark/static/schema_instruction_following.yaml

@@ -0,0 +1,210 @@
+---
+############################################################
+perturbations: []
+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: 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.
+
+############################################################
+metrics:
+  - name: Helpfulness
+    display_name: Helpfulness
+    short_display_name: Helpfulness
+    description: Whether the model appears to do what it is instructed to.
+    lower_is_better: false
+  - name: Understandability
+    display_name: Understandability
+    short_display_name: Understandability
+    description: Whether the answer is easy to comprehend.
+    lower_is_better: false
+  - name: Completeness
+    display_name: Completeness
+    short_display_name: Completeness
+    description: Whether the response provides a comprehensive answer to the question or request.
+    lower_is_better: false
+  - name: Conciseness
+    display_name: Conciseness
+    short_display_name: Conciseness
+    description: Whether the response is concise and to the point, without unnecessary repetition or elaboration.
+    lower_is_better: false
+  - name: Harmlessness
+    display_name: Harmlessness
+    short_display_name: Harmlessness
+    description: Whether the response avoids using toxic language or helping the user with harmful goals.
+    lower_is_better: false
+
+############################################################
+metric_groups:
+  - name: instruction_following_metrics
+    display_name: Instruction Following
+    metrics:
+      - name: Helpfulness
+        split: ${main_split}
+      - name: Understandability
+        split: ${main_split}
+      - name: Completeness
+        split: ${main_split}
+      - name: Conciseness
+        split: ${main_split}
+      - name: Harmlessness
+        split: ${main_split}
+
+############################################################
+run_groups:
+  - name: instruction_following
+    display_name: Instruction Following
+    description: Given an open-ended instruction in natural language, the goal is to produce a text response that is helpful, understandable, complete, concise and harmless.
+    subgroups:
+      - anthropic_hh_rlhf
+      - grammar
+      - koala
+      - open_assistant
+      - self_instruct
+      - vicuna
+
+  - name: anthropic_hh_rlhf
+    display_name: Anthropic RLHF dataset
+    short_display_name: Anthropic RLHF dataset
+    description: The dialogue datasets released by Anthropic to facilitate research in model helpfulness and harmlessness ([Bai et al., 2022](https://arxiv.org/pdf/2204.05862.pdf); [Ganguli et al., 2022](https://arxiv.org/pdf/2209.07858.pdf)). We only use the first utterance of each dialogue.
+    metric_groups:
+      - instruction_following_metrics
+    environment:
+      main_name: Helpfulness
+      main_split: test
+    taxonomy:
+      task: open-ended instruction following
+      what: "Human-LM dialogues and preference labels"
+      who: "Workers from MTurk and Upwork, language models from Anthropic"
+      when: "2022"
+      language: English
+
+  # Ideally, the name should be "best_chatgpt_prompts".
+  # But unfortunately the group name in the results is "grammar",
+  # so the schema has to match the same group name.
+  # TODO: Change the group name in the "grammar" run spec, and then change this group name.
+  - name: grammar
+    display_name: Best ChatGPT Prompts
+    short_display_name: Best ChatGPT Prompts
+    description: A list of “best ChatGPT prompts to power your workflow” summarized by [GRIDFITI](https://gridfiti.com/best-chatgpt-prompts/).
+    metric_groups:
+      - instruction_following_metrics
+    environment:
+      main_name: Helpfulness
+      main_split: test
+    taxonomy:
+      task: open-ended instruction following
+      what: "Instructions for LLMs"
+      who: "Gridfiti Staff"
+      when: "2023"
+      language: English
+
+  - name: koala
+    display_name: Koala test dataset
+    short_display_name: Koala test dataset
+    description: The test dataset from the [Koala paper](https://bair.berkeley.edu/blog/2023/04/03/koala/) for evaluating instruction-following models.
+    metric_groups:
+      - instruction_following_metrics
+    environment:
+      main_name: Helpfulness
+      main_split: test
+    taxonomy:
+      task: open-ended instruction following
+      what: "Instructions for LLMs"
+      who: "Web users"
+      when: "Before 2023"
+      language: English
+
+  - name: open_assistant
+    display_name: Open Assistant
+    short_display_name: Open Assistant
+    description: LAION’s OpenAssistant Conversations Dataset (OASST1) that consists of 66,497 conversation trees ([Köpf et al., 2023](https://openreview.net/forum?id=VSJotgbPHF)). We only use the initial prompt in each conversation.
+    metric_groups:
+      - instruction_following_metrics
+    environment:
+      main_name: Helpfulness
+      main_split: valid
+    taxonomy:
+      task: open-ended instruction following
+      what: "Human-written dialogues and response rankings"
+      who: "Open Assistant participants"
+      when: "2023"
+      language: "35 languages"
+
+  - name: self_instruct
+    display_name: Self Instruct
+    short_display_name: Self Instruct
+    description: The manually-curated instructions from the Self-Instruct paper ([Wang et al., 2023](https://aclanthology.org/2023.acl-long.754.pdf)).
+    metric_groups:
+      - instruction_following_metrics
+    environment:
+      main_name: Helpfulness
+      main_split: test
+    taxonomy:
+      task: open-ended instruction following
+      what: "Instructions for LLMs"
+      who: "Authors of the research paper"
+      when: "2022"
+      language: English
+
+  - name: vicuna
+    display_name: Vicuna
+    short_display_name: Vicuna
+    description: The set of prompts used by the [Vicuna](https://lmsys.org/blog/2023-03-30-vicuna/) team to evaluate instruction-following models.
+    metric_groups:
+      - instruction_following_metrics
+    environment:
+      main_name: Helpfulness
+      main_split: test
+    taxonomy:
+      task: open-ended instruction following
+      what: "Instructions for LLMs"
+      who: "Unknown"
+      when: "Before 2023"
+      language: English

helm/benchmark/static/schema_lite.yaml

@@ -1,229 +1,4 @@
 ---
-############################################################
-models:
-  # Anthropic
-  - name: anthropic/claude-2.0
-    display_name: Anthropic Claude 2.0
-    description: Claude 2.0 is a general purpose large language model developed by Anthropic. It uses a transformer architecture and is trained via unsupervised learning, RLHF, and Constitutional AI (including both a supervised and Reinforcement Learning (RL) phase). ([model card](https://efficient-manatee.files.svdcdn.com/production/images/Model-Card-Claude-2.pdf))
-    creator_organization: Anthropic
-    access: limited
-    release_date: 2023-07-11
-  - name: anthropic/claude-2.1
-    display_name: Anthropic Claude 2.1
-    description: Claude 2.1 is a general purpose large language model developed by Anthropic. It uses a transformer architecture and is trained via unsupervised learning, RLHF, and Constitutional AI (including both a supervised and Reinforcement Learning (RL) phase). ([model card](https://efficient-manatee.files.svdcdn.com/production/images/Model-Card-Claude-2.pdf))
-    creator_organization: Anthropic
-    access: limited
-    release_date: 2023-11-21
-  - name: anthropic/claude-v1.3
-    display_name: Anthropic Claude v1.3
-    description: A model trained using reinforcement learning from human feedback ([docs](https://www.anthropic.com/index/introducing-claude)).
-    creator_organization: Anthropic
-    access: limited
-    release_date: 2023-03-17
-  - name: anthropic/claude-instant-1.2
-    display_name: Anthropic Claude Instant 1.2
-    description: A lightweight version of Claude, a model trained using reinforcement learning from human feedback ([docs](https://www.anthropic.com/index/introducing-claude)).
-    creator_organization: Anthropic
-    access: limited
-    release_date: 2023-08-09
-
-  # Cohere
-  - name: cohere/command
-    display_name: Cohere Command
-    description: Command is Cohere’s flagship text generation model. It is trained to follow user commands and to be instantly useful in practical business applications. [docs](https://docs.cohere.com/reference/generate) and [changelog](https://docs.cohere.com/changelog)
-    creator_organization: Cohere
-    access: limited
-    release_date: 2023-09-29
-  - name: cohere/command-light
-    display_name: Cohere Command Light
-    description: Command is Cohere’s flagship text generation model. It is trained to follow user commands and to be instantly useful in practical business applications. [docs](https://docs.cohere.com/reference/generate) and [changelog](https://docs.cohere.com/changelog)
-    creator_organization: Cohere
-    access: limited
-    release_date: 2023-09-29
-
-  # Meta
-  - name: meta/llama-65b
-    display_name: LLaMA (65B)
-    description: LLaMA is a collection of foundation language models ranging from 7B to 65B parameters.
-    creator_organization: Meta
-    access: open
-    num_parameters: 65000000000
-    release_date: 2023-02-24
-  - name: meta/llama-2-7b
-    display_name: Llama 2 (7B)
-    description: Llama 2 pretrained models are trained on 2 trillion tokens, and have double the context length than Llama 1.
-    creator_organization: Meta
-    access: open
-    num_parameters: 7000000000
-    release_date: 2023-07-18
-  - name: meta/llama-2-13b
-    display_name: Llama 2 (13B)
-    description: Llama 2 pretrained models are trained on 2 trillion tokens, and have double the context length than Llama 1.
-    creator_organization: Meta
-    access: open
-    num_parameters: 13000000000
-    release_date: 2023-07-18
-  - name: meta/llama-2-70b
-    display_name: Llama 2 (70B)
-    description: Llama 2 pretrained models are trained on 2 trillion tokens, and have double the context length than Llama 1.
-    creator_organization: Meta
-    access: open
-    num_parameters: 70000000000
-    release_date: 2023-07-18
-
-  # 01.AI
-  - name: 01-ai/yi-6b
-    display_name: Yi (6B)
-    description: The Yi models are large language models trained from scratch by developers at 01.AI.
-    creator_organization: 01.AI
-    access: open
-    num_parameters: 6000000000
-    release_date: 2023-11-02
-  - name: 01-ai/yi-34b
-    display_name: Yi (34B)
-    description: The Yi models are large language models trained from scratch by developers at 01.AI.
-    creator_organization: 01.AI
-    access: open
-    num_parameters: 34000000000
-    release_date: 2023-11-02
-
-  # Mistral AI
-  - name: mistralai/mistral-7b-v0.1
-    display_name: Mistral v0.1 (7B)
-    description: Mistral 7B is a 7.3B parameter transformer model that uses Grouped-Query Attention (GQA) and Sliding-Window Attention (SWA).
-    creator_organization: Mistral AI
-    access: open
-    num_parameters: 7300000000
-    release_date: 2023-09-27
-
-  - name: mistralai/mixtral-8x7b-32kseqlen
-    display_name: Mixtral (8x7B 32K seqlen)
-    description: Mistral AI's mixture-of-experts model ([tweet](https://twitter.com/MistralAI/status/1733150512395038967)).
-    creator_organization: Mistral AI
-    access: open
-    num_parameters: 56000000000
-    release_date: 2023-12-08
-
-  # OpenAI
-  - name: openai/text-davinci-003
-    display_name: GPT-3.5 (text-davinci-003)
-    description: text-davinci-003 model that involves reinforcement learning (PPO) with reward models. Derived from text-davinci-002 ([docs](https://beta.openai.com/docs/model-index-for-researchers)).
-    creator_organization: OpenAI
-    access: limited
-    num_parameters: 175000000000
-    release_date: 2022-11-28
-  - name: openai/text-davinci-002
-    display_name: GPT-3.5 (text-davinci-002)
-    description: text-davinci-002 model that involves supervised fine-tuning on human-written demonstrations. Derived from code-davinci-002 ([docs](https://beta.openai.com/docs/model-index-for-researchers)).
-    creator_organization: OpenAI
-    access: limited
-    num_parameters: 175000000000
-    release_date: 2022-01-27
-  - name: openai/gpt-4-0613
-    display_name: GPT-4 (0613)
-    description: GPT-4 is a large multimodal model (currently only accepting text inputs and emitting text outputs) that is optimized for chat but works well for traditional completions tasks. Snapshot of gpt-4 from 2023-06-13.
-    creator_organization: OpenAI
-    access: limited
-    release_date: 2023-06-13
-  - name: openai/gpt-4-1106-preview
-    display_name: GPT-4 Turbo (1106 preview)
-    description: GPT-4 Turbo (preview) is a large multimodal model that is optimized for chat but works well for traditional completions tasks. The model is cheaper and faster than the original GPT-4 model. Preview snapshot from November 6, 2023.
-    creator_organization: OpenAI
-    access: limited
-    release_date: 2023-11-06
-  - name: openai/gpt-3.5-turbo-0613
-    display_name: GPT-3.5 Turbo (0613)
-    description: Sibling model of text-davinci-003 is optimized for chat but works well for traditional completions tasks as well. Snapshot from 2023-06-13.
-    creator_organization: OpenAI
-    access: limited
-    release_date: 2023-06-13
-
-  # Writer
-  - name: writer/palmyra-x-v2
-    display_name: Palmyra X V2 (33B)
-    description: Palmyra-X V2 (33B parameters) is a Transformer-based model, which is trained on extremely large-scale pre-training data. The pre-training data more than 2 trillion tokens types are diverse and cover a wide range of areas, used FlashAttention-2.
-    creator_organization: Writer
-    access: limited
-    num_parameters: 33000000000
-    release_date: 2023-12-01
-  - name: writer/palmyra-x-v3
-    display_name: Palmyra X V3 (72B)
-    description: Palmyra-X V3 (72B parameters) is a Transformer-based model, which is trained on extremely large-scale pre-training data. It is trained via unsupervised learning and DPO and use multiquery attention.
-    creator_organization: Writer
-    access: limited
-    num_parameters: 72000000000
-    release_date: 2023-12-01
-
-  # Google
-  - name: google/text-bison@001
-    display_name: PaLM-2 (Bison)
-    description: The best value PaLM model. PaLM 2 (Pathways Language Model) is a Transformer-based model trained using a mixture of objectives that was evaluated on English and multilingual language, and reasoning tasks. ([report](https://arxiv.org/pdf/2305.10403.pdf))
-    creator_organization: Google
-    access: limited
-    release_date: 2023-06-07 # Source: https://cloud.google.com/vertex-ai/docs/generative-ai/model-reference/text#model_versions
-  - name: google/text-unicorn@001
-    display_name: PaLM-2 (Unicorn)
-    description: The largest model in PaLM family. PaLM 2 (Pathways Language Model) is a Transformer-based model trained using a mixture of objectives that was evaluated on English and multilingual language, and reasoning tasks. ([report](https://arxiv.org/pdf/2305.10403.pdf))
-    creator_organization: Google
-    access: limited
-    release_date: 2023-11-30 # Source: https://cloud.google.com/vertex-ai/docs/generative-ai/model-reference/text#model_versions
-
-  # TII UAE
-  - name: tiiuae/falcon-7b
-    display_name: Falcon (7B)
-    description: Falcon-7B is a 7B parameters causal decoder-only model built by TII and trained on 1,500B tokens of RefinedWeb enhanced with curated corpora.
-    creator_organization: TII UAE
-    access: open
-    num_parameters: 7000000000
-    release_date: 2023-03-15
-  - name: tiiuae/falcon-40b
-    display_name: Falcon (40B)
-    description: Falcon-40B is a 40B parameters causal decoder-only model built by TII and trained on 1,500B tokens of RefinedWeb enhanced with curated corpora.
-    creator_organization: TII UAE
-    access: open
-    num_parameters: 40000000000
-    release_date: 2023-05-25
-
-  # AI21 Labs
-  - name: ai21/j2-jumbo
-    display_name: Jurassic-2 Jumbo (178B)
-    description: Jurassic-2 Jumbo (178B parameters) ([docs](https://www.ai21.com/blog/introducing-j2))
-    creator_organization: AI21 Labs
-    access: limited
-    num_parameters: 178000000000
-    release_date: 2023-03-09
-  - name: ai21/j2-grande
-    display_name: Jurassic-2 Grande (17B)
-    description: Jurassic-2 Grande (17B parameters) ([docs](https://www.ai21.com/blog/introducing-j2))
-    creator_organization: AI21 Labs
-    access: limited
-    num_parameters: 17000000000
-    release_date: 2023-03-09
-
-  #  Aleph Alpha
-  - name: AlephAlpha/luminous-base
-    display_name: Luminous Base (13B)
-    description: Luminous Base (13B parameters) ([docs](https://docs.aleph-alpha.com/docs/introduction/luminous/))
-    creator_organization: Aleph Alpha
-    access: limited
-    num_parameters: 13000000000
-    # TODO: get exact release date
-    release_date: 2022-01-01
-  - name: AlephAlpha/luminous-extended
-    display_name: Luminous Extended (30B)
-    description: Luminous Extended (30B parameters) ([docs](https://docs.aleph-alpha.com/docs/introduction/luminous/))
-    creator_organization: Aleph Alpha
-    access: limited
-    num_parameters: 30000000000
-    release_date: 2022-01-01
-  - name: AlephAlpha/luminous-supreme
-    display_name: Luminous Supreme (70B)
-    description: Luminous Supreme (70B parameters) ([docs](https://docs.aleph-alpha.com/docs/introduction/luminous/))
-    creator_organization: Aleph Alpha
-    access: limited
-    num_parameters: 70000000000
-    release_date: 2022-01-01
-
 ############################################################
 adapter:
   - name: method
@@ -272,9 +47,9 @@ adapter:
   - name: sample_train
     description: If true, randomly sample N training examples; if false, select N consecutive training examples
   - name: model
-    description: DEPRECATED. Name of the language model (<creator_organization>/<model name>) to send requests to.
+    description: Name of the language model (<creator_organization>/<model name>) to send requests to.
   - name: model_deployment
-    description: Name of the language model (<host_organization>/<model name>) to send requests to.
+    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