persistent-fm/cpt-3.6B

CTD Pretrain Corpus — v1 (capped)

Pretraining corpus assembled from clinical-trial, regulatory, biomedical-literature, code, SQL, and synthetic-generator sources. Every record is a JSON object with a text field (plus source-specific metadata/source fields).

This is the v1 / capped release: each source category was streamed in file order and truncated once its configured token budget was reached (whole-record cutoff — the record that crosses the limit is kept, so totals land at or just over the stated cap). Categories with no configured limit (generators) are included in full. The uncapped/full version of every category lives in the sibling processed-all release.

Token counts were computed with the local tokenizer at overall/v1/qwen3.5-0.8b (Qwen 3.5 0.8B tokenizer), counting the text field only, with add_special_tokens=False.

Overall stats

Category	Tokens	Records	Token limit	Limit reached	Files
clinical_trials	1,000,000,017	222,238	1,000,000,000	yes	clinical_trials.jsonl
generators	892,358,213	705,472	— (no limit)	no	t0_llm.jsonl, t1.jsonl, t2_dedup.jsonl, t3_q_dedup.jsonl, t4_kpis_v2_relations.jsonl, t5_templated.jsonl, t7.jsonl
generic	500,000,092	725,848	500,000,000	yes	generic.jsonl
python	400,000,261	486,395	400,000,000	yes	python.jsonl
sql	400,009,402	54,065	400,000,000	yes	sql.jsonl
fda	150,000,137	1,694,490	150,000,000	yes	fda.jsonl
pmc	300,022,934	29,776	300,000,000	yes	pmc.jsonl
Total	3,642,391,056	3,918,284

Generators — per-file breakdown

The generators category has no token cap, so every source file is included in full. Individual file stats:

File	Records	Tokens
t0_llm.jsonl	12,473	13,544,109
t1.jsonl	637,765	765,998,078
t2_dedup.jsonl	30,998	6,123,972
t3_q_dedup.jsonl	9,839	92,449,908
t4_kpis_v2_relations.jsonl	92	1,542,592
t5_templated.jsonl	14,184	12,266,811
t7.jsonl	121	432,743
Subtotal (generators)	705,472	892,358,213

Source categories

• clinical_trials — clinical trial records (from ct-v1/ct-v1.jsonl), single merged file.
• generators — synthetic/templated data-generation outputs (t0–t7 pipeline stages), kept as individual files, no cap applied.
• generic — general-domain text sampled from a ~750M-token pool (generic/all-750M/*.parquet), merged into one file.
• python — Python/code + visualization datasets (python/*.parquet), merged into one file.
• sql — SQL Q&A / schema data (sql/*.arrow), merged into one file.
• fda — FDA regulatory pretraining text (fda_pretrain_full.jsonl).
• pmc — PubMed Central biomedical literature (pmc.jsonl).

Format

Each file is newline-delimited JSON (.jsonl). Every record contains at least:

{"text": "...", "metadata": {...}}

metadata/source fields vary by category and reflect the original upstream schema.

Tokenizer

• Path: overall/v1/qwen3.5-0.8b
• Counting method: Tokenizer.encode_batch(texts, add_special_tokens=False), token count = len(encoding.ids) per record's text field.

Files in this release

See stats.json / stats.md in this directory for the machine-readable and tabular versions of the stats above.