evals-for-every-language / system_architecture_diagram.md
davidpomerenke's picture
Upload from GitHub Actions: updated frontend and backend to fix bugs
4e8cb1a verified
# AI Language Monitor - System Architecture
This diagram shows the complete data flow from model discovery through evaluation to frontend visualization.
```mermaid
flowchart TD
%% Model Sources
A1["important_models<br/>Static Curated List"] --> D[load_models]
A2["get_historical_popular_models<br/>Web Scraping - Top 20"] --> D
A3["get_current_popular_models<br/>Web Scraping - Top 10"] --> D
A4["blocklist<br/>Exclusions"] --> D
%% Model Processing
D --> |"Combine & Dedupe"| E["Dynamic Model List<br/>~40-50 models"]
E --> |get_or_metadata| F["OpenRouter API<br/>Model Metadata"]
F --> |get_hf_metadata| G["HuggingFace API<br/>Model Details"]
G --> H["Enriched Model DataFrame"]
H --> |Save| I[models.json]
%% Model Validation & Cost Filtering
H --> |"Validate Models<br/>Check API Availability"| H1["Valid Models Only<br/>Cost ≀ $20/1M tokens"]
H1 --> |"Timeout Protection<br/>120s for Large Models"| H2["Robust Model List"]
%% Language Data
J["languages.py<br/>BCP-47 + Population"] --> K["Top 100 Languages"]
%% Task Registry with Unified Prompting
L["tasks.py<br/>7 Evaluation Tasks"] --> M["Task Functions<br/>Unified English Zero-Shot"]
M --> M1["translation_from/to<br/>BLEU + ChrF"]
M --> M2["classification<br/>Accuracy"]
M --> M3["mmlu<br/>Accuracy"]
M --> M4["arc<br/>Accuracy"]
M --> M5["truthfulqa<br/>Accuracy"]
M --> M6["mgsm<br/>Accuracy"]
%% On-the-fly Translation with Origin Tagging
subgraph OTF [On-the-fly Dataset Translation]
direction LR
DS_raw["Raw English Dataset<br/>"] --> Google_Translate["Google Translate API"]
Google_Translate --> DS_translated["Translated Dataset<br/>(e.g., MGSM/ARC)<br/>Origin: 'machine'"]
DS_native["Native Dataset<br/>(e.g., AfriMMLU/Global-MMLU)<br/>Origin: 'human'"]
end
%% Evaluation Pipeline
H2 --> |"models ID"| N["main.py / main_gcs.py<br/>evaluate"]
K --> |"languages bcp_47"| N
L --> |"tasks.items"| N
N --> |"Filter by model.tasks"| O["Valid Combinations<br/>Model Γ— Language Γ— Task"]
O --> |"10 samples each"| P["Evaluation Execution<br/>Batch Processing"]
%% Task Execution with Origin Tracking
P --> Q1[translate_and_evaluate<br/>Origin: 'human']
P --> Q2[classify_and_evaluate<br/>Origin: 'human']
P --> Q3[mmlu_and_evaluate<br/>Origin: 'human' (no on-the-fly for missing; uses auto-translated dataset if available)]
P --> Q4[arc_and_evaluate<br/>Origin: 'human'/'machine']
P --> Q5[truthfulqa_and_evaluate<br/>Origin: 'human' (no on-the-fly for missing; relies on available datasets)]
P --> Q6[mgsm_and_evaluate<br/>Origin: 'human'/'machine']
%% API Calls with Error Handling
Q1 --> |"complete() API<br/>Rate Limiting"| R["OpenRouter<br/>Model Inference"]
Q2 --> |"complete() API<br/>Rate Limiting"| R
Q3 --> |"complete() API<br/>Rate Limiting"| R
Q4 --> |"complete() API<br/>Rate Limiting"| R
Q5 --> |"complete() API<br/>Rate Limiting"| R
Q6 --> |"complete() API<br/>Rate Limiting"| R
%% Results Processing with Origin Aggregation
R --> |Scores| S["Result Aggregation<br/>Mean by model+lang+task+origin"]
S --> |Save| T[results.json]
%% Backend & Frontend with Origin-Specific Metrics
T --> |Read| U[backend.py]
I --> |Read| U
U --> |make_model_table| V["Model Rankings<br/>Origin-Specific Metrics"]
U --> |make_country_table| W["Country Aggregation"]
U --> |"API Endpoint"| X["FastAPI /api/data<br/>arc_accuracy_human<br/>arc_accuracy_machine"]
X --> |"JSON Response"| Y["Frontend React App"]
%% UI Components
Y --> Z1["WorldMap.js<br/>Country Visualization"]
Y --> Z2["ModelTable.js<br/>Model Rankings"]
Y --> Z3["LanguageTable.js<br/>Language Coverage"]
Y --> Z4["DatasetTable.js<br/>Task Performance"]
%% Data Sources with Origin Information
subgraph DS ["Data Sources"]
DS1["Flores-200<br/>Translation Sentences<br/>Origin: 'human'"]
DS2["MMLU/AfriMMLU/Global-MMLU<br/>Knowledge QA<br/>Origin: 'human' or 'machine' (HF auto-translated only)"]
DS3["ARC<br/>Science Reasoning<br/>Origin: 'human'"]
DS4["TruthfulQA<br/>Truthfulness<br/>Origin: 'human'"]
DS5["MGSM<br/>Math Problems<br/>Origin: 'human'"]
end
DS1 --> Q1
DS2 --> Q3
DS3 --> Q4
DS4 --> Q5
DS5 --> Q6
%% No on-the-fly DS_translated for MMLU anymore; only HF auto-translated used
DS_translated --> Q4
DS_translated --> Q5
DS_native --> Q3
DS_native --> Q4
DS_native --> Q5
%% Styling - Neutral colors that work in both dark and light modes
classDef modelSource fill:#f8f9fa,stroke:#6c757d,color:#212529
classDef evaluation fill:#e9ecef,stroke:#495057,color:#212529
classDef api fill:#dee2e6,stroke:#6c757d,color:#212529
classDef storage fill:#d1ecf1,stroke:#0c5460,color:#0c5460
classDef frontend fill:#f8d7da,stroke:#721c24,color:#721c24
classDef translation fill:#d4edda,stroke:#155724,color:#155724
class A1,A2,A3,A4 modelSource
class Q1,Q2,Q3,Q4,Q5,Q6,P evaluation
class R,F,G,X api
class T,I storage
class Y,Z1,Z2,Z3,Z4 frontend
class Google_Translate,DS_translated,DS_native translation
```
## Architecture Components
### πŸ”΅ Model Discovery (Light Gray)
- **Static Curated Models**: Handpicked important models for comprehensive evaluation
- **Dynamic Popular Models**: Real-time discovery of trending models via web scraping
- **Quality Control**: Blocklist for problematic or incompatible models
- **Model Validation**: API availability checks and cost filtering (≀$20/1M tokens)
- **Timeout Protection**: 120s timeout for large/reasoning models, 60s for others
- **Metadata Enrichment**: Rich model information from OpenRouter and HuggingFace APIs
### 🟣 Evaluation Pipeline (Medium Gray)
- **7 Active Tasks**: Translation (bidirectional), Classification, MMLU, ARC, TruthfulQA, MGSM
- **Unified English Zero-Shot Prompting**: All tasks use English instructions with target language content
- **Origin Tagging**: Distinguishes between human-translated ('human') and machine-translated ('machine') data
- **Combinatorial Approach**: Systematic evaluation across Model Γ— Language Γ— Task combinations
- **Sample-based**: 10 evaluations per combination for statistical reliability
- **Batch Processing**: 50 tasks per batch with rate limiting and error resilience
- **Dual Deployment**: `main.py` for local/GitHub, `main_gcs.py` for Google Cloud with GCS storage
### 🟠 API Integration (Light Gray)
- **OpenRouter**: Primary model inference API for all language model tasks
- **Rate Limiting**: Intelligent batching and delays to prevent API overload
- **Error Handling**: Graceful handling of timeouts, rate limits, and model unavailability
- **HuggingFace**: Model metadata and open-source model information
- **Google Translate**: Specialized translation API for on-the-fly dataset translation
### 🟒 Data Storage (Cyan)
- **results.json**: Aggregated evaluation scores with origin-specific metrics
- **models.json**: Dynamic model list with metadata and validation status
- **languages.json**: Language information with population data
### 🟑 Frontend Visualization (Light Red)
- **WorldMap**: Interactive country-level language proficiency visualization
- **ModelTable**: Ranked model performance leaderboard with origin-specific columns
- **LanguageTable**: Language coverage and speaker statistics
- **DatasetTable**: Task-specific performance breakdowns with human/machine distinction
### πŸ”΅ Translation & Origin Tracking (Light Green)
- **On-the-fly Translation**: Google Translate API for languages without native benchmarks
- **Origin Tagging**: Automatic classification of data sources (human vs. machine translated)
- **Separate Metrics**: Frontend displays distinct scores for human and machine-translated data
## Data Flow Summary
1. **Model Discovery**: Combine curated + trending models β†’ validate API availability β†’ enrich with metadata
2. **Evaluation Setup**: Generate all valid Model Γ— Language Γ— Task combinations with origin tracking
3. **Task Execution**: Run evaluations using unified English prompting and appropriate datasets
4. **Result Processing**: Aggregate scores by model+language+task+origin and save to JSON files
5. **Backend Serving**: FastAPI serves processed data with origin-specific metrics via REST API
6. **Frontend Display**: React app visualizes data through interactive components with transparency indicators
This architecture enables scalable, automated evaluation of AI language models across diverse languages and tasks while providing real-time insights through an intuitive web interface with methodological transparency.