Remek/basal-1.5-max
Remek/basal-1.5-max is an 11 billion parameter model from the basal-1.5 family, developed by Remek. It is designed as a dynamic classifier for typed decisions, providing calibrated probabilities for predefined answers rather than generating free text. This model excels at tasks requiring precise classification, such as document routing, policy checks, RAG decisions, and risk scoring, supporting both Polish and English. It is the largest and most accurate model in its family, optimized for long documents and complex decision-making.
Loading preview...
basal-1.5-max: A Typed-Decision Model for Classification
basal-1.5-max is an 11 billion parameter model developed by Remek, part of the basal-1.5 family. Unlike traditional LLMs that generate free text, this model functions as a dynamic classifier by providing calibrated probabilities for predefined, typed answers (e.g., choice, yes/no, score). It processes a given 'state' (document, message, agent trace) and answers questions about it in a single forward pass, ensuring answers always fall within specified options.
Key Capabilities
- Dynamic Classification: Classes are described in plain language within the request, allowing one model to serve diverse tasks without retraining.
- Calibrated Probabilities: Returns a confidence score for each possible answer, indicating the model's certainty.
- Multilingual Support: Optimized for both Polish and English decision-making.
- High Accuracy: Achieves 0.773 macro accuracy on the Werdykt benchmark (Polish 0.777, English 0.769) and 0.933 on Polish sealed tests.
- Versatile Deployment: Supports various inference engines including its native basal engine (PyTorch), vLLM, SGLang, MLX (Apple Silicon), Ollama, and llama.cpp.
- Confidence-Based Automation: Provides confidence thresholds for 1% and 5% error targets, enabling automated decision-making for a significant portion of tasks (e.g., 62.2% coverage at 1% error).
Ideal Use Cases
- Document Routing: Directing inquiries or documents to appropriate departments based on content.
- Policy and Rule Checks: Verifying compliance against defined rules or policies.
- RAG Decisions: Assessing relevance, sufficiency, conflicts, or claim support in Retrieval-Augmented Generation workflows.
- Urgency and Risk Scoring: Assigning quantitative scores for triage and prioritization.
- Judging Answers: Evaluating the correctness or quality of generated responses.
- Agent and Tool Decisions: Guiding the actions of AI agents or tool selection.
- Long Document Analysis: Excels in making decisions based on extensive textual inputs.