Add ResearchRadar project

.github/workflows/staging_run_complete_llm.yml

@@ -45,6 +45,18 @@ jobs:
         run: |
           zenml init
 
+      - name: List and describe ZenML projects
+        working-directory: ./llm-complete-guide
+        run: |
+          zenml project list || echo "Could not list projects"
+          zenml project describe || echo "Failed to describe project"
+
+      - name: Register Set ZenML project
+        working-directory: ./llm-complete-guide
+        run: |
+          zenml project register llm-complete-guide || echo "Failed to register project"
+          zenml project set llm-complete-guide || echo "Failed to set project"
+
       - name: Set stack (Staging)
         working-directory: ./llm-complete-guide
         run: |

.typos.toml

@@ -5,6 +5,7 @@ extend-exclude = [
     "*.ipynb",
     "llm-finetuning/*",
     "end-to-end-computer-vision/*",
+    "research-radar/data/*.jsonl",
 ]
 
 [default.extend-identifiers]
@@ -17,6 +18,8 @@ MDEyOk9yZ2FuaXphdGlvbjg4Njc2OTU1 = "MDEyOk9yZ2FuaXphdGlvbjg4Njc2OTU1"
 
 [default.extend-words]
 # Don't correct the surname "Teh"
+preprocesser = "preprocesser"
+Preprocesser = "Preprocesser"
 aks = "aks"
 hashi = "hashi"
 womens = "womens"
@@ -39,6 +42,12 @@ colours = "colours"
 initialised = "initialised"
 visualisation = "visualisation"
 customise = "customise"
+modernbert = "modernbert"
+deepseek = "deepseek"
+huggingface = "huggingface"
+answerdotai = "answerdotai"
+preprocessor = "preprocessor"
+logits = "logits"
 
 [default]
 locale = "en-us"

README.md

@@ -72,6 +72,7 @@ A list of updated and maintained projects by the ZenML team and the community:
 | [Eurorate Predictor](eurorate-predictor) | `ETL` `Time Series` `Feature Engineering` `Regression` `Workflow Orchestration` | `cloud-composer` `airflow` `vertex-ai` `bigquery` `xgboost` `gcp` |
 | [Nightwatch AI](nightwatch-ai) | `NLP` `Text Summarization` `Database Integration` `LLM` `Automated Reporting` | `openai` `supabase` `slack` `github-actions` `gcp` |
 | [Sign Language Detection with YOLOv5](sign-language-detection-yolov5) | `Computer Vision` `Object Detection` `Real-time Processing` `Model Deployment` | `mlflow` `gcp` `bentoml` `vertex-ai` `docker` |
+| [ResearchRadar](research-radar) | `AI Literature Discovery` `Research Paper Classification` `Model Training` `Model Evaluation` `Model Comparison` | `anthropic` `huggingface` `pytorch` `transformers` `docker` |
 
 # 💻 System Requirements
 

llmops-article-classifier/README.md → research-radar/README.md

@@ -1,11 +1,21 @@
-# 🚀 LLMOps Article Classifier: Automated End-to-End ML Pipelines with ZenML
+<div align="center">
+  <img src="assets/training_pipeline.png" alt="ResearchRadar Training Pipeline DAG" width="800"/>
+</div>
+
+# 🚀 ResearchRadar: AI Literature Discovery and Classification Pipeline
 
 ## 📑 Overview
 
-This project demonstrates ZenML's capabilities for building end-to-end ML pipelines that automatically classify articles for the ZenML LLMOps Database. It showcases a complete MLOps workflow that:
+This project demonstrates how AI researchers can build an automated pipeline to discover, classify, and organize research papers relevant to their specific subfields. Using ZenML's end-to-end ML pipelines, researchers can train models to automatically identify papers that align with their research interests—whether it's reinforcement learning, computer vision, NLP, or emerging areas like multimodal models—from across various sources including preprint servers, journals, and conference proceedings.
+
+The system showcases how researchers can fine-tune models to recognize subtle distinctions between related subfields, saving countless hours normally spent manually filtering through the overwhelming volume of AI research being published daily. The comparative analysis pipeline demonstrates whether custom-trained models outperform commercial APIs when identifying highly specialized technical content, helping research teams make informed decisions about their literature review infrastructure.
+
+This framework is ideal for academic labs, industry research groups, and individual AI researchers who need to maintain comprehensive awareness of relevant publications without spending hours each week manually sifting through papers. By deploying the trained model to Hugging Face, teams can create shared resources that help all members stay current with literature relevant to their collective research agenda, streamlining collaboration and reducing duplicate effort in literature reviews.
+
+This specific example showcases how to build an end-to-end ML pipeline that automatically classifies LLMOps-related research papers for the ZenML LLMOps Database. It showcases a complete MLOps workflow that:
 
 - Classifies articles using DeepSeek R1 (Classification Pipeline)
-- Fine-tunes a ModernBERT model for the classification task (Training Pipeline)
+- Fine-tunes [ModernBERT](https://www.answer.ai/posts/2024-12-19-modernbert.html) for the classification task (Training Pipeline)
 - Provides comprehensive metrics tracking (Training Pipeline)
 - Compares the fine-tuned model against Claude Haiku (Comparison Pipeline)
 - Deploys the fine-tuned model to Hugging Face (Deployment Pipeline)
@@ -107,44 +117,29 @@ python run.py deploy
 
 ### Prerequisites
 
-- Python 3.9+
-- ZenML 0.75.0+
+- Python 3.11+
+- ZenML 0.80.0+
 - Hugging Face account with API token
 - Anthropic API key (optional, for model comparison)
 
 ### Installation
 
-1. Clone the repository
-
-   ```bash
-   git clone https://github.com/zenml-io/llmops-article-classifier-demo
-   cd llmops-article-classifier-demo
-   ```
-
-2. Create and activate virtual environment
-
-   ```bash
-   python -m venv venv
-   source venv/bin/activate  # On Windows: venv\Scripts\activate
-   ```
-
-3. Install dependencies
+1. Install LLMops Article Classifier
 
-   ```bash
-   pip install -r requirements.txt
-   ```
+```bash
+# Set up a Python virtual environment
+python3 -m venv venv
+source venv/bin/activate  # On Windows: venv\Scripts\activate
 
-4. Set up environment variables
+# Install requirements
+pip install -r requirements.txt
+```
 
-   ```bash
-   cp .env.example .env
-   # Edit .env file to add your HF_TOKEN and ANTHROPIC_API_KEY
-   ```
+2. Set up environment variables
+```bash
+cp .env.example .env  # add your HF_TOKEN and ANTHROPIC_API_KEY
+```
 
-5. Initialize ZenML (if not already done)
-   ```bash
-   zenml init
-   ```
 
 ### Configuration
 
@@ -230,6 +225,18 @@ The project follows the recommended ZenML project structure:
 └── run.py              # CLI entrypoint
 ```
 
+## 📚 Additional Documentation
+
+The project includes detailed documentation in various subdirectories:
+- **[Data Documentation](data/README.md)**: Details on dataset storage and processing.
+- **[Classification Results Documentation](classification_results/README.md)**: Explanation of classification outputs, metrics, and the checkpoint system.
+- **[Model Comparison Metrics Documentation](model_compare_metrics/README.md)**: Details on the model comparison.
+- **[Pipelines Documentation](pipelines/README.md)**: Details on the pipeline definitions.
+- **[Prompts Documentation](prompts/README.md)**: Details on the prompts used in the pipeline.
+- **[Schemas Documentation](schemas/README.md)**: Details on data models and validation.
+- **[Steps Documentation](steps/README.md)**: Information about individual pipeline components including model comparison and metrics.
+- **[Utils Documentation](utils/README.md)**: Details on helper functions, checkpoint system, classification utilities, and data processing.
+
 ## 📜 License
 
 This project is licensed under the Apache License 2.0 - see the [LICENSE](LICENSE) file for details.

llmops-article-classifier/classification_results/README.md → research-radar/classification_results/README.md

@@ -58,15 +58,14 @@ A human-readable markdown report of the metrics data with formatted tables for:
 
 ## Checkpoint System
 
-The classification process uses a checkpoint system to enable resuming interrupted jobs:
+The classification process uses a checkpoint system (enabled by default) to resume interrupted jobs. 
 
 - Checkpoints are saved periodically during processing in the `checkpoints/` directory
-- The frequency and retention settings are configurable in `settings.yaml`
-- Only the most recent N checkpoints are kept to conserve disk space
+- The frequency and retention settings are configurable in [`base_config.yaml`](../configs/base_config.yaml)
 
 ## Processing Configuration
 
-Classification processing can be configured in [`settings.yaml`](../settings.yaml):
+Classification processing can be configured in [`base_config.yaml`](../configs/base_config.yaml):
 
 - Process entire dataset or specific batches
 - Enable/disable parallel processing

llmops-article-classifier/classification_results/compute_stats.py → research-radar/classification_results/compute_stats.py

@@ -50,7 +50,9 @@ def convert_column_oriented_to_list(data: dict) -> List[dict]:
         raise ValueError("Data is not in a column-oriented format.")
 
     # Assume all columns share the same sorted row indices.
-    row_indices = sorted(next(iter(data.values())).keys(), key=lambda x: int(x))
+    row_indices = sorted(
+        next(iter(data.values())).keys(), key=lambda x: int(x)
+    )
     rows = []
     for idx in row_indices:
         row = {col: data[col].get(idx) for col in columns}
@@ -210,7 +212,9 @@ def print_statistics(stats: Dict[int, Dict[str, Any]]) -> None:
         )
 
 
-def print_run_summary(run_summary: Dict[str, Any], avg_consistency: float) -> None:
+def print_run_summary(
+    run_summary: Dict[str, Any], avg_consistency: float
+) -> None:
     """
     Print overall run-level summary statistics.
 

llmops-article-classifier/configs/base_config.yaml → research-radar/configs/base_config.yaml

@@ -94,7 +94,7 @@ steps:
       source_type: "artifact" # or "disk"
       path: "artifacts/test_set" # used if source_type is "disk"
       artifact_name: "test_set"
-      version: 20
+      version: # (optional) set to specific version if source_type is "artifact"
     batch_sizes:
       modernbert: 25
       claude: 10

llmops-article-classifier/materializers/__init__.py → research-radar/materializers/__init__.py

@@ -16,9 +16,9 @@
 #
 
 # materializers/__init__.py
-from .lazyframe_materializer import LazyFrameMaterializer
-from .datetime_materializer import DatetimeMaterializer
 from .dataset_materializer import DatasetMaterializer
+from .datetime_materializer import DatetimeMaterializer
+from .lazyframe_materializer import LazyFrameMaterializer
 from .url_materializer import UrlMaterializer
 
 __all__ = [

llmops-article-classifier/materializers/dataset_materializer.py → research-radar/materializers/dataset_materializer.py

@@ -17,9 +17,10 @@
 
 # materializers/dataset_materializer.py
 from typing import Type
+
 from datasets import Dataset, load_from_disk
-from zenml.materializers.base_materializer import BaseMaterializer
 from zenml.enums import ArtifactType
+from zenml.materializers.base_materializer import BaseMaterializer
 
 
 class DatasetMaterializer(BaseMaterializer):