# GiacomoSaccaggi/scomp_link

Source: https://glimind.com/server/github%2Fgiacomosaccaggi%2Fscomp_link
Registry id: `github/giacomosaccaggi/scomp_link`
Status: unknown

## Tools (42)

- `check_fairness`: Compute fairness metrics: demographic parity, disparate impact (4/5 rule), equalized odds. Returns whether the model is fair and detailed group-level metrics.
- `check_version`: Check the current scomp-link version and whether an update is available. Returns: - installed_version: Currently installed version - pypi_version: Latest versio
- `cluster_data`: Cluster data using KMeans or MeanShift. Returns cluster labels and silhouette score. Optionally saves result with cluster column.
- `compare_models`: Compare multiple .scomp artifacts side by side. artifacts is a comma-separated list of paths. Returns a comparison table of metrics.
- `create_visualization`: Create a single visualization. chart_type can be any of the 39 available charts. Categories: plotly (histogram, barchart, linechart, area_chart), rawgraphs (tre
- `describe_data`: Profile a dataset. Returns column-level stats: dtype, missing%, unique count, min, max, mean, std. Use this first to understand any new dataset before training 
- `detect_anomalies`: Detect anomalies using multi-method consensus (Isolation Forest, LOF, TabNet, Transformer). Returns number of anomalies and per-method comparison.
- `detect_drift`: Detect distribution drift between reference (training) and current (production) data. Returns PSI per feature and which features drifted.
- `embed_text`: Generate embeddings from a trained contrastive text model (.scomp artifact). Returns embedding shape and optionally saves to .npy file. Use this to extract text
- `engineer_features`: Apply automated feature engineering: polynomial interactions, log transforms for skewed features, date extraction, target encoding. Returns output path and new 
- `export_model`: Export a .scomp model to standard format: pickle, joblib, onnx. Returns the output file path and size.
- `forecast_series`: Forecast future values of a time series column. Methods: auto, arima, exp_smoothing. Returns predicted values and optional plot.
- `generate_report`: Generate an interactive HTML report. If only data is provided, creates an EDA report. If artifact is also provided, creates a model evaluation report with predi
- `llm_convert`: Convert a HuggingFace model to quantized GGUF format for efficient inference with llama.cpp. Parameters: model_path: Path to the HuggingFace model directory. qu
- `llm_dedup`: Deduplicate a text dataset using exact hashing or MinHash near-duplicate detection. Parameters: data: Path to text file (one text per line) or JSON/JSONL file. 
- `llm_estimate`: Estimate the GGUF output size for a model and quantization level without performing the conversion. Parameters: model_path: Path to the HuggingFace model direct
- `llm_evaluate`: Evaluate generated text quality using n-gram metrics (BLEU, ROUGE, diversity, Zipf). Parameters: generated: Generated text or path to a text file. references: R
- `llm_finetune`: Fine-tune a pretrained HuggingFace model using LoRA, QLoRA, or full fine-tuning. Parameters: model: HuggingFace model identifier or local path. data: Path to th
- `llm_format`: Convert an instruction-tuning dataset between formats (Alpaca, ShareGPT, OpenAI → ChatML, Llama, Plain). Parameters: input_path: Path to input JSON or JSONL fil
- `llm_merge`: Merge multiple fine-tuned models using Linear, SLERP, TIES, or DARE strategies. Parameters: base_model: Path to base model (shared pretrained ancestor). models:
- `llm_scratch`: Build and train a GPT-style transformer model from scratch. Parameters: data: Path to the training corpus. vocab_size: Vocabulary size. d_model: Model embedding
- `llm_serve`: Get info about serving a model (does NOT start the server — use the Python API for that). Parameters: model_path: Path to HuggingFace model directory. port: Por
- `predict`: Generate predictions from a .scomp artifact on new data. Returns predictions array. Optionally saves to CSV.
- `report_add_chart`
- `report_add_code`
- `report_add_comparison_table`
- `report_add_dark_mode_toggle`: Add a floating dark/light mode toggle button to the report. Inserts a fixed-position button in the top-right corner that switches between light mode and dark mo
- `report_add_diff`
- `report_add_kpi_cards`
- `report_add_mermaid`
- `report_add_section`: Open a new collapsible section in the report. Automatically closes the previous section if one is open. Parameters: report_id: The report ID returned by report_
- `report_add_summary_stats`: Add an auto-generated data profiling summary table to the report. Displays a compact table with column statistics: type, non-null count, missing %, unique value
- `report_add_table`: Add a data table to the report. The table is interactive (sortable, CSV-exportable). Parameters: report_id: The report ID returned by report_create. data: JSON 
- `report_add_tabs`
- `report_add_terminal`
- `report_add_text`: Add text content to the current section of a report. Parameters: report_id: The report ID returned by report_create. content: The text or HTML content to add. s
- `report_create`
- `report_save`: Save the report to an HTML file and close the session. Parameters: report_id: The report ID returned by report_create. output: Output file path (e.g. "my_report
- `select_backbone`: Find the best pretrained embedding backbone for a dataset. Evaluates multiple sentence-transformer models by computing contrastive loss on a sample. Returns ran
- `train_model`: Train an ML model. Supports regression, classification, text, clustering. Set engineer=true for automatic feature engineering. Set tune=true for Optuna hyperpar
- `update_scomp_link`: Update scomp-link to the latest version from PyPI. Parameters: force: If true, update even if already on latest version or checked recently. Returns status with
- `validate_model`: Evaluate a saved model on test data. Returns metrics (regression: r2, rmse, mae; classification: accuracy, f1, precision, recall). Optionally generates an HTML 

Measured by Glimind (https://glimind.com): MCP handshake + tool list only, no tool is ever called. CC BY 4.0.
