v4: coherent labels (37K) — Claude-distilled extractive compression, argmax inference

v3: trained on 330K structured tool outputs (H100) — JSON, diffs, logs, code, SQL, agentic traces

v2: trained on 15K labels (3x more data) from agentic + function-calling + coding traces

Initial release: Kompress v1 — ModernBERT token compressor for agentic contexts

initial commit