Datasets:

QuixiAI
/

SYN-1B

language:
  - en
license: apache-2.0
tags:
  - synthetic
  - pretraining
  - text-generation
  - language-modeling
task_categories:
  - text-generation
task_ids:
  - language-modeling
size_categories:
  - 1M<n<10M
pretty_name: SYN-1B
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet
      - split: validation
        path: data/validation-*.parquet
train-eval-index:
  - config: default
    task: text-generation
    task_id: language-modeling
    splits:
      train_split: train
      eval_split: validation
    col_mapping:
      text: text

SYN-1B

Dataset Summary

SYN-1B is a 1.04B-token synthetic language-modeling corpus of rule-governed text streams. Each row is a decoded instance in which the text establishes facts, mappings, bindings, or simple generative rules; later spans may revise those rules, swap bindings, delay a query over long filler, or present a null control with event-like surface text that should not change the answer.

The dataset is intended as structured pretraining data, not as a supervised instruction dataset. It uses ordinary next-token prediction, but the examples are generated from known latent state machines. That makes it useful when a researcher wants training data where long-context behavior, variable binding, belief revision, recency, distractor robustness, and held-out transfer can be measured exactly rather than inferred from natural text.

The data/ Parquet files expose one decoded synthetic instance per row in the text column. The original packed uint16 token shards, sidecars, and QA manifests are kept under raw/ for exact token-level training and audit use. The Hugging Face validation split corresponds to the SYN generator's eval split. F5 appears only in that held-out eval/validation split.

Why Pretrain With It?

SYN-1B gives language models repeated practice on behaviors that are important in real use but sparse, ambiguous, or hard to label in web text:

Tracking a rule or binding across many intervening tokens.
Updating an answer after a correction instead of copying the most recent local surface pattern.
Preserving unchanged facts when only part of a rule changes.
Distinguishing real state-changing events from plausible-looking distractors.
Answering from evidence that may be hundreds or thousands of tokens old.
Transferring the same abstract update pattern to a held-out surface format.

For general pretraining, SYN-1B can be mixed as a small synthetic fraction alongside web, education, code, or domain corpora to add dense, auditable examples of state tracking and revision. It is especially relevant for work on long-context models, recurrent or memory-augmented models, retrieval-free reasoning, continual belief updating, and curricula that target systematic generalization.

Other Applications

Beyond pretraining, SYN-1B can be used as:

A controlled benchmark for long-context recall and rule updating.
A source of probing data for whether hidden states encode active mappings, bindings, recency, and corrected values.
A stress test for distractor sensitivity and surface-marker shortcuts.
A data generator baseline for synthetic-curriculum, data-mixture, and scaling-law studies.
A reproducible corpus for mechanistic interpretability, because the raw sidecars under raw/sidecars/ identify events, writes, query positions, answer positions, evidence spans, composition depth, controlled gap lengths, and held-out families.

Limitations

SYN-1B is synthetic by design. It should not be treated as a substitute for natural-language pretraining data, an instruction-following dataset, or a factuality benchmark. The task families cover a controlled set of rule-updating and long-context patterns, so improvements on SYN-1B should be checked against naturalistic and out-of-distribution evaluations before making broad claims.

The decoded Parquet view is convenient for Hugging Face tooling, while the packed token shards and sidecars in raw/ are the reference artifacts for exact token-level audits and reproduction. Users who need full supervision metadata should read the sidecars rather than relying only on the compact Parquet columns.

Task Families

The task families isolate different forms of state tracking:

Family	Role
F1 branch reversal	A rule holds, then an event changes the active mapping.
F2 binding swap	Entity-to-attribute bindings are corrected by swaps rather than a single surface marker.
F3 delayed correction / recall	Old evidence and controlled evidence age test whether stored evidence survives and can be reinterpreted.
F4 flat null	No interpretive event occurs; event-like distractors test false-positive rule updates.
F5 modular-stream switch	Eval-only transfer task with no training exposure, used to test transfer beyond memorized surface formats.

Source and Reproducibility

The generator, QA suite, and build specification are available in the source repository. See SYN-1B.md for the corpus specification.

How to Load

from datasets import load_dataset

ds = load_dataset("QuixiAI/SYN-1B")

Dataset Stats

Split	Rows	Synthetic instance tokens in raw build
`train`	4,063,978	1,000,001,463
`validation`	178,863	40,002,882
`total`	4,242,841	1,040,004,345

The Hugging Face dataset is decoded text. Consumers can tokenize the text column with any tokenizer. The token counts above refer to the reference raw build in raw/, which was generated and QA-audited with the HuggingFaceTB/SmolLM2-135M tokenizer.

Data Schema

Columns:

text: decoded synthetic instance text.
instance_id, family, syn_split: stable identifiers and SYN family.
token_len, task_token_count, filler_token_count, controlled_gap_tokens, composition_depth, num_queries: compact per-instance metadata.
shard, window_index, start_offset, token_hash: join keys back to the packed raw token stream and sidecar records under raw/.

Licensing Information

License: Apache-2.0