allenai/state-bench

state-tracking-long-context-v1

Synthetic evaluations of state tracking across four algebraic families: r-trivial, aperiodic, solvable, and periodic. Every family is derived from one swap-based periodic source trace: solvable rewrites each copy as a modular self-increment (x <- (x + d) mod n) landing on the copied value, aperiodic as the copied literal, and r-trivial as the target's first value.

Available surface formatters: integer-code-modular, spreadsheet-cells-modular, cube-painting, status-lights, people-in-rooms.

Each rendered record contains prefix, completion, stable example_id and semantic instance_id fields, four incorrect_completions, shuffled choices, a zero-based correct_choice_index, its complexity class, and generation parameters. The canonical config contains formatter-independent operation traces and final states. See generation.yaml, formats/, and manifest.json for exact reproduction metadata.

Rendered configs are partitioned into tokenizer-length splits named <split>__<stratum> using allenai/Olmo-3-1025-7B over prefix. The half-open strata are tokens_under_2k=[0, 2048), tokens_4k=[2,048, 4096), tokens_8k=[4,096, 8192), tokens_16k=[8,192, 16384), tokens_32k=[16,384, 32768), tokens_64k=[32,768, 65536), tokens_128k=[65,536, 131072), tokens_256k=[131,072, 262144), tokens_512k=[262,144, 524288), tokens_1m=[524,288, 1048576), tokens_2m_plus=[1,048,576, infinity). Each row records its token_length and token_stratum; token lengths are measured by exact per-row tokenization. The token_length_is_estimate field distinguishes sampled estimates from exact measurements. Empty tokenizer ranges are omitted, so consumers should discover the available split names for each rendered config. The canonical config retains the unstratified base splits.

Reference assignments use eager copy semantics: assigning one variable from another copies its value at that point in the trace and does not create an alias. Increments wrap modulo the condition's value count.