Datasets:

domofon
/

structured-cpt

license: mit
language:
  - en
task_categories:
  - text-generation
tags:
  - pretrain
  - continued-pretraining
  - structured
  - sql
  - json
  - smollm2
  - domofon
pretty_name: Structured CPT (JSON + SQL)
size_categories:
  - 1B<n<10B
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/*.parquet

Structured CPT - JSON + SQL pretrain documents

SmolLM2-1.7B continued-pretraining shard of structured documents. Each document is a <task> / <input> / <output> block whose <output> is a canonical JSON object, terminated by the SmolLM2 end-of-text token ``.

Sources:

source	description	rows	shards	repeat
`sql_bmc2`	b-mc2 sql-create-context -> JSON (4 keys, stub explanation)	392,885	1	5
`sql_gretelai`	gretelai synthetic_text_to_sql -> JSON (4 keys)	529,255	1	5
`sql_hardikch`	hardikch05 100000_text_to_sql -> JSON (question/schema_ddl/sql)	5,500,000	28	1
`stackexchange_json`	Stack Exchange -> JSON (title/body/tags)	836,000	5	1

Totals: 7,258,140 rows.

from datasets import load_dataset
ds = load_dataset("domofon/structured-cpt", split="train")
print(ds[0]["text"])

Document shape

<task>
...
</task>

<input>
...
</input>

<output>
{...canonical json...}
</output>

The JSON inside <output> keeps the documented key order and exact (JSON-escaped) source values. SQL values are copied verbatim from the input.

Columns

column	description
`id`	unique id `{source}-{index}`
`text`	full structured document, trailing ``
`text_preview`	first 800 characters of `text`
`n_chars`	length of `text`
`source`	one of stackexchange_json / sql_gretelai / sql_bmc2 / sql_hardikch

Shards

35 parquet files under data/ (zstd-compressed).

Identity

Built for the Domofon CPT stack; see sibling datasets domofon/worldsim-domofon-cpt and domofon/domofon-identity-anchor.