WIP: initial version of checks, creating a Gradio UI for the app

app.py

@@ -1,28 +1,54 @@
 import gradio as gr
-from huggingface_hub import ModelCard
+
 from compliance_checks import (
     ComplianceSuite,
     ModelProviderIdentityCheck,
-    IntendedPurposeCheck
+    IntendedPurposeCheck,
+    GeneralLimitationsCheck,
+    ComputationalRequirementsCheck,
 )
 
+from bloom_card import bloom_card
 
-def run_compliance_check(repo_name):
-    model_card = ModelCard.load(repo_id_or_path=repo_name).content
 
+def run_compliance_check(model_card: str):
     suite = ComplianceSuite(checks=[
         ModelProviderIdentityCheck(),
-        IntendedPurposeCheck()
+        IntendedPurposeCheck(),
+        GeneralLimitationsCheck(),
+        ComputationalRequirementsCheck(),
     ])
 
     results = suite.run(model_card)
 
-    return str(results)
+    return str([r[0] for r in results])
+
+
+with gr.Blocks() as demo:
+    gr.Markdown("""\
+    # Model Card Validator
+    Following Article 13 of the EU AI Act
+    """)
+
+    with gr.Row():
+        with gr.Column():
+            model_card_box = gr.TextArea()
+            populate_sample = gr.Button(value="Populate Sample")
+            submit = gr.Button()
+
+        with gr.Column():
+            results_list = gr.Text()
+
+    submit.click(
+        fn=run_compliance_check,
+        inputs=[model_card_box],
+        outputs=[results_list]
+    )
 
+    populate_sample.click(
+        fn=lambda: bloom_card,
+        inputs=[],
+        outputs=[model_card_box]
+    )
 
-gr.Interface(
-    fn=run_compliance_check,
-    inputs="text",
-    outputs="text",
-    examples=[["society-ethics/model-card-webhook-test"]]
-).launch()
+demo.launch()

bloom_card.py

@@ -0,0 +1,147 @@
+bloom_card = """\
+# Model Details  
+
+BLOOM is an autoregressive Large Language Model (LLM), trained to continue text from a prompt on vast amounts of text data using industrial-scale computational resources. As such, it is able to output coherent text in 46 languages and 13 programming languages that is hardly distinguishable from text written by humans. BLOOM can also be instructed to perform text tasks it hasn't been explicitly trained for, by casting them as text generation tasks.
+
+## Basics
+*This section provides information about the model type, version, license, funders, release date, developers, and contact information.*
+*It is useful for anyone who wants to reference the model.*
+  
+**Developed by:** BigScience ([website](https://bigscience.huggingface.co))
+
+*All collaborators are either volunteers or have an agreement with their employer. (Further breakdown of participants forthcoming.)*
+    
+**Model Type:** Transformer-based Language Model
+
+**Checkpoints format:** `transformers` (Megatron-DeepSpeed format available [here](https://huggingface.co/bigscience/bloom-optimizer-states))
+
+**Version:** 1.0.0
+
+**Languages:** Multiple; see [training data](#training-data)
+
+**License:** RAIL License v1.0 ([link](https://huggingface.co/spaces/bigscience/license) / [article and FAQ](https://bigscience.huggingface.co/blog/the-bigscience-rail-license))
+
+**Release Date Estimate:** Monday, 11.July.2022
+
+**Send Questions to:** [email protected]
+
+**Cite as:** BigScience, _BigScience Language Open-science Open-access Multilingual (BLOOM) Language Model_. International, May 2021-May 2022
+
+**Funded by:** 
+    
+* The French government.
+
+* Hugging Face ([website](https://huggingface.co)).
+
+* Organizations of contributors.  *(Further breakdown of organizations forthcoming.)*
+
+## Intended Use
+
+This model is being created in order to enable public research on large language models (LLMs). LLMs are intended to be used for language generation or as a pretrained base model that can be further fine-tuned for specific tasks. Use cases below are not exhaustive.
+
+### Direct Use
+
+-   Text generation
+
+-   Exploring characteristics of language generated by a language model
+
+    -   Examples: Cloze tests, counterfactuals, generations with reframings
+
+### Downstream Use
+
+-   Tasks that leverage language models include: Information Extraction, Question Answering, Summarization
+
+### Out-of-Scope Use
+
+Using the model in [high-stakes](#high-stakes) settings is out of scope for this model.  The model is not designed for [critical decisions](#critical-decisions) nor uses with any material consequences on an individual's livelihood or wellbeing. The model outputs content that appears factual but may not be correct.  
+
+Out-of-scope Uses Include:
+
+-   Usage in biomedical domains, political and legal domains, or finance domains
+
+-   Usage for evaluating or scoring individuals, such as for employment, education, or credit
+
+-   Applying the model for critical automatic decisions, generating factual content, creating reliable summaries, or generating predictions that must be correct
+
+#### Misuse
+
+Intentionally using the model for harm, violating [human rights](#human-rights), or other kinds of malicious activities, is a misuse of this model. This includes:
+
+-   Spam generation
+
+-   Disinformation and influence operations
+
+-   Disparagement and defamation
+
+-   Harassment and abuse
+  
+-   [Deception](#deception)
+
+-   Unconsented impersonation and imitation
+
+-   Unconsented surveillance 
+
+-   Generating content without attribution to the model, as specified in the [RAIL License, Use Restrictions](https://huggingface.co/spaces/bigscience/license)
+
+## Bias, Risks, and Limitations
+*This section identifies foreseeable harms and misunderstandings.*
+    
+Model may:
+
+-   Overrepresent some viewpoints and underrepresent others
+
+-   Contain stereotypes
+  
+-   Contain [personal information](#personal-data-and-information)
+
+-   Generate:
+
+    -   Hateful, abusive, or violent language
+
+    -   Discriminatory or prejudicial language
+
+    -   Content that may not be appropriate for all settings, including sexual content
+
+-   Make errors, including producing incorrect information as if it were factual
+
+-   Generate irrelevant or repetitive outputs
+
+-   Induce users into attributing human traits to it, such as sentience or consciousness
+
+## Technical Specifications
+*This section includes details about the model objective and architecture, and the compute infrastructure.*
+*It is useful for people interested in model development.*
+
+### Compute infrastructure
+Jean Zay Public Supercomputer, provided by the French government (see [announcement](https://www.enseignementsup-recherche.gouv.fr/fr/signature-du-marche-d-acquisition-de-l-un-des-supercalculateurs-les-plus-puissants-d-europe-46733)).
+
+#### Hardware
+
+* 384 A100 80GB GPUs (48 nodes)
+    
+* Additional 32 A100 80GB GPUs (4 nodes) in reserve
+
+* 8 GPUs per node Using NVLink 4 inter-gpu connects, 4 OmniPath links
+
+* CPU: AMD
+
+* CPU memory: 512GB per node
+
+* GPU memory: 640GB per node
+
+* Inter-node connect: Omni-Path Architecture (OPA)
+
+* NCCL-communications network: a fully dedicated subnet
+
+* Disc IO network: shared network with other types of nodes
+
+#### Software
+
+* Megatron-DeepSpeed ([Github link](https://github.com/bigscience-workshop/Megatron-DeepSpeed))
+
+* DeepSpeed ([Github link](https://github.com/microsoft/DeepSpeed))
+
+* PyTorch (pytorch-1.11 w/ CUDA-11.5; see [Github link](https://github.com/pytorch/pytorch))
+
+* apex ([Github link](https://github.com/NVIDIA/apex))
+"""

compliance_checks.py

@@ -6,16 +6,16 @@ from bs4 import BeautifulSoup, Comment
 
 class ComplianceCheck(ABC):
     @abstractmethod
-    def run_check(self, card: BeautifulSoup) -> bool:
+    def run_check(self, card: BeautifulSoup):
         raise NotImplementedError
 
 
 class ModelProviderIdentityCheck(ComplianceCheck):
     def run_check(self, card: BeautifulSoup):
         try:
-            model_description = card.find("h3", string="Model Description")
-            description_list = model_description.find_next_siblings()[0]
-            developer = description_list.find(string="Developed by:").parent.next_sibling.strip()
+            developed_by = card.find("strong", string="Developed by:")
+
+            developer = "".join([str(s) for s in developed_by.next_siblings]).strip()
 
             if developer == "[More Information Needed]":
                 return False, None
@@ -26,6 +26,8 @@ class ModelProviderIdentityCheck(ComplianceCheck):
 
 
 def walk_to_next_heading(card, heading, heading_text):
+    stop_at = [heading, f"h{int(heading[1]) - 1}"]
+
     try:
         heading_node = card.find(heading, string=heading_text)
 
@@ -34,10 +36,10 @@ def walk_to_next_heading(card, heading, heading_text):
         sibling_gen = heading_node.nextSiblingGenerator()
         sibling = next(sibling_gen)
 
-        while not (sibling.name is not None and sibling.name.startswith("h")) or sibling.name is None:
+        while sibling and (not (sibling.name is not None and sibling.name in stop_at) or sibling.name is None):
             if not isinstance(sibling, Comment):
                 content = content + sibling.text.strip()
-            sibling = next(sibling_gen)
+            sibling = next(sibling_gen, None)
 
         if content.strip() == "[More Information Needed]":
             return False, None
@@ -50,6 +52,7 @@ def walk_to_next_heading(card, heading, heading_text):
 class IntendedPurposeCheck(ComplianceCheck):
     def run_check(self, card: BeautifulSoup):
         direct_use_check, direct_use_content = walk_to_next_heading(card, "h3", "Direct Use")
+        # TODO: Handle [optional], which doesn't exist in BLOOM, e.g.
         downstream_use_check, downstream_use_content = walk_to_next_heading(card, "h3", "Downstream Use [optional]")
         out_of_scope_use_check, out_of_scope_use_content = walk_to_next_heading(card, "h3", "Out-of-Scope Use")
         return (
@@ -58,6 +61,16 @@ class IntendedPurposeCheck(ComplianceCheck):
         )
 
 
+class GeneralLimitationsCheck(ComplianceCheck):
+    def run_check(self, card: BeautifulSoup):
+        return walk_to_next_heading(card, "h2", "Bias, Risks, and Limitations")
+
+
+class ComputationalRequirementsCheck(ComplianceCheck):
+    def run_check(self, card: BeautifulSoup):
+        return walk_to_next_heading(card, "h3", "Compute infrastructure")
+
+
 class ComplianceSuite:
     def __init__(self, checks):
         self.checks = checks

main.py

@@ -1,78 +0,0 @@
-import os
-
-from huggingface_hub import (comment_discussion,
-                             create_discussion, get_discussion_details,
-                             get_repo_discussions)
-from tabulate import tabulate
-from difflib import SequenceMatcher
-
-KEY = os.environ.get("KEY")
-
-
-def similar(a, b):
-    """Check similarity of two sequences"""
-    return SequenceMatcher(None, a, b).ratio()
-
-
-def create_metadata_breakdown_table(compliance_check_dictionary):
-    data = {k: v for k, v in compliance_check_dictionary.items()}
-    metadata_fields_column = list(data.keys())
-    metadata_values_column = list(data.values())
-    table_data = list(zip(metadata_fields_column, metadata_values_column))
-    return tabulate(
-        table_data, tablefmt="github", headers=("Compliance Check", "Present")
-    )
-
-
-def create_markdown_report(
-    desired_metadata_dictionary, repo_name, update: bool = False
-):
-    report = f"""# Model Card Regulatory Compliance report card {"(updated)" if update else ""}
-    \n
-This is an automatically produced model card regulatory compliance report card for {repo_name}.
-This report is meant as a POC!
-    \n 
-## Breakdown of metadata fields for your model
-\n
-{create_metadata_breakdown_table(desired_metadata_dictionary)}
-\n
-    """
-    return report
-
-
-def create_or_update_report(compliance_check, repo_name):
-    report = create_markdown_report(
-        compliance_check, repo_name, update=False
-    )
-    repo_discussions = get_repo_discussions(
-        repo_name,
-        repo_type="model",
-    )
-    for discussion in repo_discussions:
-        if (
-            discussion.title == "Metadata Report Card" and discussion.status == "open"
-        ):  # An existing open report card thread
-            discussion_details = get_discussion_details(
-                repo_name, discussion.num, repo_type="model"
-            )
-            last_comment = discussion_details.events[-1].content
-            if similar(report, last_comment) <= 0.999:
-                report = create_markdown_report(
-                    compliance_check,
-                    repo_name,
-                    update=True,
-                )
-                comment_discussion(
-                    repo_name,
-                    discussion.num,
-                    comment=report,
-                    repo_type="model",
-                )
-            return True
-    create_discussion(
-        repo_name,
-        "Model Card Regulatory Compliance Report Card",
-        description=report,
-        repo_type="model",
-    )
-    return True

server.py

@@ -1,25 +0,0 @@
-import os
-from fastapi import FastAPI, Request, Response
-from main import parse_webhook_post, run_compliance_check, create_or_update_report
-
-KEY = os.environ.get("KEY")
-
-app = FastAPI()
-
-
-@app.post("/webhook")
-async def webhook(request: Request):
-    if request.method == "POST":
-        # if request.headers.get("X-Webhook-Secret") != KEY:
-        #     return Response("Invalid secret", status_code=401)
-
-        data = await request.json()
-
-        if parsed_post := parse_webhook_post(data):
-            repo_name = parsed_post
-        else:
-            return Response("Unable to parse webhook data", status_code=400)
-
-        compliance_check = run_compliance_check(repo_name)
-        result = create_or_update_report(compliance_check, repo_name)
-        return result

tests/conftest.py

@@ -0,0 +1,7 @@
+import pytest
+from bloom_card import bloom_card as bc
+
+@pytest.fixture()
+def bloom_card():
+    # TODO: Note, this is a heavily doctored version of the card.
+    return bc

tests/test_compliance_checks.py

@@ -2,8 +2,23 @@ import pytest
 from unittest.mock import MagicMock
 
 import markdown
-from bs4 import BeautifulSoup, Comment
-from compliance_checks import ComplianceSuite, ModelProviderIdentityCheck, IntendedPurposeCheck
+from bs4 import BeautifulSoup
+from compliance_checks import (
+    ComplianceSuite,
+    ModelProviderIdentityCheck,
+    IntendedPurposeCheck,
+    GeneralLimitationsCheck,
+    ComputationalRequirementsCheck,
+)
+
+
+expected_infrastructure = """\
+Jean Zay Public Supercomputer, provided by the French government.\
+Hardware\
+384 A100 80GB GPUs (48 nodes)\
+Software\
+Megatron-DeepSpeed (Github link)\
+"""
 
 
 class TestComplianceCheck:
@@ -109,11 +124,92 @@ Some random info...
 [More Information Needed]
         """
 
+    @pytest.fixture
+    def general_limitations_model_card(self):
+        return """
+# Model Card for Sample Model
+
+## Some Random Header
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+Hello world! These are some risks...
+
+## More Things
+        """
+
+    @pytest.fixture
+    def bad_general_limitations_model_card(self):
+        return """
+# Model Card for Sample Model
+
+## Some Random Header
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+## More Things
+        """
+
+    @pytest.fixture
+    def computational_requirements_model_card(self):
+        # Adapted from: https://huggingface.co/bigscience/bloom/blob/main/README.md
+        return """
+# Model Card for Sample Model
+
+## Some Random Header
+
+## Technical Specifications
+
+### Compute infrastructure
+Jean Zay Public Supercomputer, provided by the French government.
+
+#### Hardware
+
+* 384 A100 80GB GPUs (48 nodes)
+
+#### Software
+
+* Megatron-DeepSpeed ([Github link](https://github.com/bigscience-workshop/Megatron-DeepSpeed))
+</details>
+
+## Intended Use
+
+Etc..
+"""
+
+    @pytest.fixture
+    def bad_computational_requirements_model_card(self):
+        # Adapted from: https://huggingface.co/bigscience/bloom/blob/main/README.md
+        return """
+# Model Card for Sample Model
+
+## Some Random Header
+
+## Technical Specifications
+
+### Compute infrastructure
+[More Information Needed]
+
+## Intended Use
+
+Etc..
+"""
+
     @pytest.mark.parametrize("check, card,check_passed,values", [
         (ModelProviderIdentityCheck(), "provider_identity_model_card", True, "Nima Boscarino"),
         (ModelProviderIdentityCheck(), "bad_provider_identity_model_card", False, None),
         (IntendedPurposeCheck(), "intended_purpose_model_card", True, ["Here is some info about direct uses...", None, "Here is some info about out-of-scope uses..."]),
         (IntendedPurposeCheck(), "bad_intended_purpose_model_card", False, [None, None, None]),
+        (GeneralLimitationsCheck(), "general_limitations_model_card", True, "Hello world! These are some risks..."),
+        (GeneralLimitationsCheck(), "bad_general_limitations_model_card", False, None),
+        (ComputationalRequirementsCheck(), "computational_requirements_model_card", True, expected_infrastructure),
+        (ComputationalRequirementsCheck(), "bad_computational_requirements_model_card", False, None),
     ])
     def test_run_model_provider_identity_check(self, check, card, check_passed, values, request):
         card = request.getfixturevalue(card)
@@ -175,7 +271,7 @@ class TestComplianceSuite:
 
 
 class TestEndToEnd:
-    @pytest.mark.parametrize("card", [
+    @pytest.mark.parametrize("card,fixture", [
         ("""
 # Model Card for Sample Model
 
@@ -187,12 +283,54 @@ Some random info...
 
 - **Developed by:** Nima Boscarino
 - **Model type:** Yada yada yada
-        """)
+
+## Uses
+
+### Direct Use
+
+Here is some info about direct uses...
+
+### Downstream Use [optional]
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+Here is some info about out-of-scope uses...
+
+## Bias, Risks, and Limitations
+
+Hello world! These are some risks...
+
+## Technical Specifications
+
+### Compute infrastructure
+Jean Zay Public Supercomputer, provided by the French government.
+
+#### Hardware
+
+* 384 A100 80GB GPUs (48 nodes)
+
+#### Software
+
+* Megatron-DeepSpeed ([Github link](https://github.com/bigscience-workshop/Megatron-DeepSpeed))
+</details>
+
+## More Things
+        """, False),
+        ("bloom_card", True)
     ])
-    def test_end_to_end_compliance_suite(self, card):
+    def test_end_to_end_compliance_suite(self, card, fixture, request):
+        if fixture:
+            card = request.getfixturevalue(card)
+
         suite = ComplianceSuite(checks=[
             ModelProviderIdentityCheck(),
             IntendedPurposeCheck(),
+            GeneralLimitationsCheck(),
+            ComputationalRequirementsCheck()
         ])
 
-        suite.run(card)
+        results = suite.run(card)
+
+        assert all([r[0] for r in results])