aabbdev/StateBench-v1

StateBench v1

310 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 · 94.78 GiB	You need high-throughput training on a compact shared VM
dense-max-100m	100M episodes · 93.45 GiB	You need maximum-density stress training on the compact register VM
grounded-max-100m	100M episodes · 96.57 GiB	You want the dense-max oracles in domain-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
)
dense_max = load_dataset(
    "aabbdev/StateBench-v1", "dense-max-100m", split="train", streaming=True
)
grounded = load_dataset(
    "aabbdev/StateBench-v1", "grounded-max-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
dense-max-100m	24,875,000 rows	125,000 rows
grounded-max-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. dense-100m and dense-max-100m are compact register-VM curricula. dense-max-100m balances 384 curriculum labels and a pairwise schedule of 32 renderer IDs and 64 transformation IDs; those IDs describe coverage scheduling, while its published prompt surface remains intentionally VM-centric.

grounded-max-100m rematerializes the same verified dense-max trajectories through 24 domain dialects: the 16 primitive domains and 8 compositions above. Its 32 document structures and 64 metamorphic surface passes are physically present in the prompts, including data formats, code-like records, logs, traces, conversations, and nested documents. This configuration targets surface transfer while preserving the exact audited answers; it 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, patch, rollback, and restore without collateral damage
Associative addressability	Resolve exact, hierarchical, indirect, and collision-prone keys
Algorithmic recurrence	Execute counters, automata, permutations, graphs, and bounded data structures
Online adaptation	Bind and remap rules inside context without updating model weights
Interference control	Protect persistent state from noise, churn, concurrent writes, and repeated reads
Transactional and distributed reasoning	Track snapshots, messages, replicas, clocks, and conflict resolution
Streaming stability	Preserve behavior across checkpoints, chunks, retries, and long horizons

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	162.01 units / 1k chars	107–120B tokens
dense-max-100m	162.88 units / 1k chars	~107–120B tokens
grounded-max-100m	160.61 units / 1k chars	~107–120B tokens

The token range was measured on 1,000 real packed rows with GPT-2 and Qwen3 tokenizers. Actual training cost depends on chat templates, truncation, packing, and the SDFT generation policy.

Integrity

• 310M 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, dense-100m-manifest.json, dense-max-100m-manifest.json, and grounded-max-100m-manifest.json.