README.md

metadata

pretty_name: StateBench v1
task_categories:
  - text-generation
language:
  - en
tags:
  - synthetic
  - state-space-models
  - recurrent-models
  - pretraining
  - sft
  - sdft
size_categories:
  - 10M<n<100M
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train/*
      - split: validation
        path: data/validation/*
  - config_name: dense-100m
    data_files:
      - split: train
        path: data/dense-100m/train/*
      - split: validation
        path: data/dense-100m/validation/*

StateBench v1

110 million verified state episodes for pretraining, SFT, and on-policy SDFT

Train models to retain, edit, address, and transform state across long execution traces.

Choose a configuration

Configuration	Scale	Use when
default	10,000,000 programs · 7.80 GiB	You need 36 explicit families and varied surfaces
dense-100m	100M episodes · 98.61 GiB	You need high-throughput state training with grounded, structurally varied prompts

All configurations contain structured prompt, completion, and teacher-only privileged context. Every target is computed by an executable reference machine, never authored by a teacher model.

Load

from datasets import load_dataset

default = load_dataset(
    "aabbdev/StateBench-v1", "default", split="train", streaming=True
)
dense = load_dataset(
    "aabbdev/StateBench-v1", "dense-100m", split="train", streaming=True
)

Configuration	Train	Validation
default	9,949,600 rows	50,400 rows
dense-100m	24,875,000 rows	125,000 rows

Train three ways

SFT

sft = dense.select_columns(["prompt", "completion"])

On-policy SDFT with TRL

sdft = dense.select_columns(["prompt", "privileged_context"])

privileged_context is visible only to the teacher. Never concatenate it into the student prompt. It contains the verified rendered target expected by TRL SDFTTrainer.

Causal pretraining

The default configuration includes a ready-to-use text column. Dense configurations avoid storing a duplicate text view; derive it while streaming:

def to_text(row):
    prompt = row["prompt"][0]["content"]
    answer = row["completion"][0]["content"]
    return {"text": f"User:\n{prompt}\n\nAssistant:\n{answer}"}

dense_pretrain = dense.map(to_text)

Coverage map

16 primitive domains

Area	Domains
Memory	Retention & lifecycle · Capacity & eviction · Addressing & aliasing · Editing & transactions
Control	Selectivity & access · Interference & concurrency · Automata & control · Algebra & reversibility
Structures	Data structures · Graphs & spatial state · Temporal streaming · Online adaptation
Systems	Distributed replication · Agents & messaging · Integrity & error control · Runtime state

8 composition domains

Area	Compositions
State systems	Transactional + temporal · Adaptive + control · Concurrent + distributed · Graph + structures
Stress systems	Robust streaming · Multi-agent state · Algebraic automata · Memory pressure

default implements 36 explicit behavioral families. The corrected dense-100m v1.4 configuration rematerializes 100M verified trajectories through 24 domain dialects: the 16 primitive domains and 8 compositions above. Its 32 parseable document structures and 64 distinct metamorphic surface passes are physically present in the prompts, including data formats, code-like records, logs, traces, conversations, and nested documents.

The dense semantic kernel trains long-horizon set, remove, additive, multiplicative, observation, and no-op transitions. Its 384 balanced families values are curriculum labels spanning domain and profile combinations, not 384 different VM mechanisms. Domain dialects diversify how transitions are presented; they do not turn each label into a full simulator of that real-world system. StateBench remains synthetic state training, not a factual natural-language corpus.

Expected model capabilities

Capability	Training pressure
Working-memory utilization	Retain independent values across long active traces
Precise state mutation	Write, overwrite, delete, add, and scale without collateral damage
Associative addressability	Resolve exact, hierarchical, indirect, and collision-prone keys
Algorithmic recurrence	Execute compact transition sequences over long horizons
In-context rule use	Apply a supplied transition legend without updating model weights
Interference control	Protect persistent state from noise, churn, concurrent writes, and repeated reads
Surface robustness	Preserve semantics across structured documents, logs, code-like records, and conversations
Streaming stability	Preserve behavior across long traces and packed independent episodes

StateBench trains state handling, not factual knowledge. It does not by itself expand a model's physical context window, add world knowledge, or guarantee general reasoning improvements outside state-intensive tasks.

Data contract

Field	Availability	Purpose
prompt	all configs	Student-visible conversational input
completion	all configs	Verified SFT target
privileged_context	all configs	Teacher-only SDFT context
text	default; derived for dense	Causal pretraining view
semantic identity	1 default; 4 dense	Deduplication and audit
family / macro / level	scalar default; lists dense	Curriculum filtering
density_units_per_1k_chars	dense configs	Semantic density control

Dense configurations at a glance

Configuration	Density	Token estimate
dense-100m	160.04 units / 1k chars	106.6B GPT-2 · 118.9B Qwen3

The token estimates were measured on 1,000 v1.4 packed rows (seed 42). Actual training cost depends on chat templates, truncation, packing, and the SDFT generation policy.

Integrity

• 110M semantic identities are unique within their respective configurations.
• Required fields, roles, answer JSON, and SDFT isolation passed full-corpus audits.
• No benchmark name, family label, or privileged answer appears in model-visible prompts.
• SHA-256 checksums cover every Parquet shard.
• The public training corpus is separate from private StateBench evaluation material.

Release manifests: dataset_manifest.json and dense-100m-manifest.json.