ZelligeAI/tessera-compressor

Overview

ZelligeAI/tessera-compressor is a specialized 1.5 billion parameter language model, fine-tuned from Qwen2.5-Coder-1.5B-Instruct, focused on compressing English reasoning text into a compact CJK/symbol register. Its primary function is to generate high-fidelity, compressed training data, notably having minted the dataset for Tessera-Preview-9B. The model operates by segmenting reasoning blocks, classifying passages, and compressing them while ensuring critical information like numbers and identifiers are preserved through a deterministic fidelity gate.

Key Capabilities

• Text Compression: Transforms English reasoning text into a shorter CJK/symbol format.
• Fidelity Gates: Employs a deterministic lexical gate to ensure novel numbers and identifiers from the original text survive compression, preventing loss of critical data.
• High Acceptance Rate: Achieved 99.0% per-passage fidelity gate success and 100% judged semantic equivalence on 103 held-out reasoning blocks.
• Efficiency: Offers a median per-passage compression ratio of 0.716 and a net corpus savings of 30.4% (after fallbacks), with downstream production use showing a median 0.58x token cost compared to English source.
• Local Inference: Designed for local deployment, eliminating the need for API keys or external dependencies for data generation.

Good For

• Training Data Generation: Ideal for creating compressed-register training data for other models, particularly for tasks involving code-centric reasoning.
• Cost-Effective Compression: Provides a low-cost alternative to frontier models for text compression, suitable for large-scale data processing.
• Preserving Key Information: Useful in scenarios where text needs to be shortened but specific numerical or identifier data must remain intact.
• Research in Compressed Reasoning: Offers a practical tool for exploring and implementing compressed reasoning techniques.