Mirror of https://github.com/roostorg/coop
github.com/roostorg/coop
Adding model card for Zentropi (#200)
* clean up integrations in enum definition to remove unused ones
* add model card details to zentropi manifest
* revert model card learn more URL to general MC educational resource
* Regenerate GraphQL files
* merge main and reinstall npm
* actually add regenerated graphql files
* revert package-lock.json to main
* include package-lock.json changes for consistent codegen
+45
-26
+1
.gitignore
+1
.gitignore
-6
client/src/graphql/generated.ts
-6
client/src/graphql/generated.ts
···
1307
1307
};
1308
1308
1309
1309
export const GQLIntegration = {
1310
-
Akismet: 'AKISMET',
1311
1310
GoogleContentSafetyApi: 'GOOGLE_CONTENT_SAFETY_API',
1312
-
L1Ght: 'L1GHT',
1313
-
MicrosoftAzureContentModerator: 'MICROSOFT_AZURE_CONTENT_MODERATOR',
1314
-
Oopspam: 'OOPSPAM',
1315
1311
OpenAi: 'OPEN_AI',
1316
-
SightEngine: 'SIGHT_ENGINE',
1317
-
TwoHat: 'TWO_HAT',
1318
1312
Zentropi: 'ZENTROPI',
1319
1313
} as const;
1320
1314
+19
package-lock.json
+19
package-lock.json
···
830
830
}
831
831
}
832
832
},
833
+
"node_modules/@graphql-codegen/cli/node_modules/@types/node": {
834
+
"version": "25.5.2",
835
+
"resolved": "https://registry.npmjs.org/@types/node/-/node-25.5.2.tgz",
836
+
"integrity": "sha512-tO4ZIRKNC+MDWV4qKVZe3Ql/woTnmHDr5JD8UI5hn2pwBrHEwOEMZK7WlNb5RKB6EoJ02gwmQS9OrjuFnZYdpg==",
837
+
"license": "MIT",
838
+
"optional": true,
839
+
"peer": true,
840
+
"dependencies": {
841
+
"undici-types": "~7.18.0"
842
+
}
843
+
},
833
844
"node_modules/@graphql-codegen/cli/node_modules/cosmiconfig": {
834
845
"version": "9.0.0",
835
846
"resolved": "https://registry.npmjs.org/cosmiconfig/-/cosmiconfig-9.0.0.tgz",
···
5260
5271
"engines": {
5261
5272
"node": ">=0.10.0"
5262
5273
}
5274
+
},
5275
+
"node_modules/undici-types": {
5276
+
"version": "7.18.2",
5277
+
"resolved": "https://registry.npmjs.org/undici-types/-/undici-types-7.18.2.tgz",
5278
+
"integrity": "sha512-AsuCzffGHJybSaRrmr5eHr81mwJU3kjw6M+uprWvCXiNeN9SOGwQ3Jn8jb8m3Z6izVgknn1R0FTCEAP2QrLY/w==",
5279
+
"license": "MIT",
5280
+
"optional": true,
5281
+
"peer": true
5263
5282
},
5264
5283
"node_modules/universalify": {
5265
5284
"version": "2.0.0",
-6
server/graphql/generated.ts
-6
server/graphql/generated.ts
···
1376
1376
};
1377
1377
1378
1378
export const GQLIntegration = {
1379
-
Akismet: 'AKISMET',
1380
1379
GoogleContentSafetyApi: 'GOOGLE_CONTENT_SAFETY_API',
1381
-
L1Ght: 'L1GHT',
1382
-
MicrosoftAzureContentModerator: 'MICROSOFT_AZURE_CONTENT_MODERATOR',
1383
-
Oopspam: 'OOPSPAM',
1384
1380
OpenAi: 'OPEN_AI',
1385
-
SightEngine: 'SIGHT_ENGINE',
1386
-
TwoHat: 'TWO_HAT',
1387
1381
Zentropi: 'ZENTROPI',
1388
1382
} as const;
1389
1383
-6
server/graphql/modules/integration.ts
-6
server/graphql/modules/integration.ts
+25
-8
server/services/integrationRegistry/integrationManifests.ts
+25
-8
server/services/integrationRegistry/integrationManifests.ts
···
158
158
159
159
const ZENTROPI: IntegrationManifestEntry = {
160
160
modelCard: {
161
-
modelName: 'Zentropi',
161
+
modelName: 'Zentropi: CoPE-A-9B',
162
162
version: '1.x',
163
-
releaseDate: 'Ongoing',
163
+
releaseDate: 'July 20, 2025',
164
164
sections: [
165
165
{
166
166
id: 'trainingData',
167
167
title: 'Training Data Sources',
168
-
fields: [{ label: 'Data Sources', value: 'TBD' }],
168
+
fields: [{ label: 'Data Sources', value: "CoPE-A's dataset includes ~60,000 labels across unique policy/content pairs using policy texts created by CoPE team and content from publicly-accessible internet forums. The CoPE team used a mix of automated and manual annotation to create golden labels. The training data includes but is not limited to hate speech, sexual content, self-harm, harassment, toxicity." }],
169
169
},
170
170
{
171
171
id: 'policyAndTaxonomy',
172
172
title: 'Policy & Taxonomy Definitions',
173
-
fields: [{ label: 'Policies', value: 'TBD' }],
173
+
fields: [{ label: 'Policies', value: "No fixed taxonomy. CoPE-A is policy-adaptive and steerable by users who define custom criteria for their specific use case. Trained for generalizable policy understanding across diverse policy formulations." }],
174
174
},
175
175
{
176
176
id: 'annotationMethodology',
177
177
title: 'Annotation Methodology',
178
-
fields: [{ label: 'Methodology', value: 'TBD' }],
178
+
fields: [{ label: 'Methodology', value: 'CoPE-A was trained using a novel training methodology that moves beyond policy memorization to achieve true policy interpretation. Trained across conflicting policy formulations with focus on generalizable policy understanding and interpretation consistency. Combined automated and manual labeling processes for quality assurance.' }],
179
179
},
180
180
{
181
181
id: 'performanceBenchmarks',
182
182
title: 'Performance Benchmarks',
183
-
fields: [{ label: 'Benchmarks', value: 'TBD' }],
183
+
fields: [{ label: 'Benchmarks', value: 'Tested on policies and content never seen during training. High accuracy across all content types: Hate Speech 91% accurate (internal test), 84% accurate (public Ethos test); Inappropriate Sexual Content 89%; Toxic Speech 90%; Self-Harm 88%; Harassment 73%. Outperforms comparable models including GPT-4o, Llama-3.1-8B, LlamaGuard3-8B, and ShieldGemma-9B across most categories.' }],
184
184
},
185
185
{
186
186
id: 'biasAndLimitations',
187
187
title: 'Bias Documentation & Known Limits',
188
-
fields: [{ label: 'Known Limitations', value: 'TBD' }],
188
+
fields: [{ label: 'Known Limitations', value: 'Text processing is limited to 8K tokens. Optimized for US English only; performance degrades for other languages/locales. Binary classification only (label present/absent). Cannot classify content requiring external verification unless explicitly defined in policy. Requires careful policy design to mitigate potential biases. Users should monitor classification patterns across demographic groups and audit decisions regularly.' }],
189
189
},
190
190
{
191
191
id: 'implementationGuidance',
···
194
194
{
195
195
label: 'Credentials',
196
196
value:
197
-
'API Key plus optional Labeler Versions (id and label per version).',
197
+
'API Key plus optional Labeler Versions (id and label per version) created by the user.',
198
+
},
199
+
{
200
+
label: 'Input Format',
201
+
value:
202
+
'Use Zentropi to create and test labeling policies and classifiers. The input format should include: (1) Overview of policy subject, (2) Definition of Terms (precise definitions of words/phrases), (3) Interpretation of Language (guidance on ambiguous language), (4) Definition of Labels with Includes/Excludes criteria. CoPE-A then returns binary classification (0/1) indicating if content matches any policy labels.',
198
203
},
199
204
],
200
205
},
···
202
207
id: 'relevantLinks',
203
208
title: 'Relevant Links',
204
209
fields: [
210
+
{
211
+
label: 'HuggingFace Model Card',
212
+
value: 'https://huggingface.co/zentropi-ai/cope-a-9b',
213
+
},
205
214
{
206
215
label: 'Documentation',
207
216
value: 'https://docs.zentropi.ai',
217
+
},
218
+
{
219
+
label: 'Research Talk',
220
+
value: 'https://www.youtube.com/live/JMq49FZ5qmY?si=Q6qpHNeTo-Bc6t9a&t=1',
221
+
},
222
+
{
223
+
label: 'Sample Code Notebook',
224
+
value: 'https://colab.research.google.com/drive/1LBmQ3d0OVrq2EpVP0tc03POalf3sDpjl?usp=sharing',
208
225
},
209
226
],
210
227
},