Pre-built OOLONG eval dataset (trec_coarse split from oolongbench/oolong-synth).

Contents: 550 rows of trec_coarse validation data across 11 context lengths (1K, 2K, 4K, 8K, 16K, 32K, 64K, 128K, 256K, 512K, 1M) — matching the RLM paper's evaluation configuration.

Format: Gzipped JSONL (validation.jsonl.gz), ~134 MB compressed, ~535 MB decompressed.

Usage: The eval harness downloads this automatically via npx tsx eval/download.ts (default --from-release mode). Use --from-hf to regenerate from HuggingFace instead.