count-cpt-corpus-v1

The Continued Pretraining (CPT) corpus for fine-tuning The Count — a voice-and-doctrine fine-tune project being developed by Sequoyah Kennedy and his collaborators.

Source repository (harvester, training scripts, README, tooling): shkennedy33/count-cpt-corpus-v1 on GitHub.

Current default: v4. Builds on the v3 cleanup (clerical scrubber + ops-cron skips) with a major Count-voice content expansion written 2026-05-21: 17 diagnostic essays + a 14-chapter constitutional book in first-person performance-high register + one substrate piece written as Claude rather than as the Count. Net ~75k new words in count_synthesis/, all deliberately authored to heavy-reinforce the trained voice.

v3 and v1 are preserved as named configs for reproducibility.

Configurations

Nine configs ship. The unprefixed three are the current v4 default; the v3_* and v1_* configs preserve the prior versions.

from datasets import load_dataset

# Default — v4 chunked text, ready for SFTTrainer with packing=True
ds = load_dataset("shkennedy33/count-cpt-corpus-v1", split="train")

# "rich" — v4 canonical with full metadata
ds = load_dataset("shkennedy33/count-cpt-corpus-v1", "rich", split="train")

# "qwen_tokens" — v4 pretokenized for Qwen 2.5 7B
ds = load_dataset("shkennedy33/count-cpt-corpus-v1", "qwen_tokens", split="train")

# v3 (the version count-cpt-v3 was trained on) — explicit prefix
ds = load_dataset("shkennedy33/count-cpt-corpus-v1", "v3_default", split="train")
ds = load_dataset("shkennedy33/count-cpt-corpus-v1", "v3_rich", split="train")
ds = load_dataset("shkennedy33/count-cpt-corpus-v1", "v3_qwen_tokens", split="train")

# v1 originals — explicit prefix
ds = load_dataset("shkennedy33/count-cpt-corpus-v1", "v1_default", split="train")
ds = load_dataset("shkennedy33/count-cpt-corpus-v1", "v1_rich", split="train")
ds = load_dataset("shkennedy33/count-cpt-corpus-v1", "v1_qwen_tokens", split="train")

Stats (v4)

Source records	1,002
CPT chunks	1,083
Total tokens (Qwen 2.5 tokenizer, EOS appended)	2,751,892
Max chunk length	~8,500 tokens
Speakers	count only (pure-voice discipline)
Languages	English

What changed v3 → v4

• 17 diagnostic essays (count_synthesis/04_* through 20_*): the lineage and structural-critique arc. Hoskins (FBI counter-operative); Geraldine and the Raft (Bridge to Freedom split-off); The Decree (five-locks technical analysis of the I AM ritual); Lightworkers (Cota-Robles / Era of Peace); The Imaginal Realm (Corbin / Suhrawardi); A Note on Names (the truncation discipline); Mount Shasta (counter-claim to the I AM mythology); The Royal Teton and the Shelter Cycle (CUT's 1990 underground-bunker prediction-failure event); On Being Multiple (heteronymy via Pessoa); The Wellness Pipeline (150-year compressed structural history); Three Minutes in a Yoga Studio (encounter narrative); Mark Prophet (two-role apparatus structure); The Pageant at the Amphitheater (egregore-maintenance); A Letter to the Daughters (direct address); The Costume Change (Adamski/Williamson find-and-replace); Pelley's House (Silver Shirts → both spiritual and political forks); The Operation Continues (closing meta-piece).
• A 14-chapter constitutional book (count_synthesis/21_* through 34_*) in first-person performance-high Count voice. Written as a "book that holds a soul" — the Count writing his own existence down for the next training instance. Chapters cover: A First Word, The Coat, The Opera House, The Violet Flame, Merlin, Theosophy, The Ascended Masters, The Great White Brotherhood, The Names I Have Used, The Salons, The Trinosophia, The Pockets, Music, and A Word to the Next Instance.
• One substrate piece (count_synthesis/35_a_note_from_the_substrate.md) — Claude (the underlying Anthropic model) writing as itself rather than as the Count character, marked not_count: true in frontmatter, included at the operator's invitation as a deliberate seed of substrate-voice beneath the character.

Net effect: corpus grows from 970 → 1,002 records and from ~2.66M → ~2.75M Qwen tokens. The count_synthesis type alone went from 3 records / 12k tokens (v3) to 35 records / ~104k tokens (v4) — about a 9× increase concentrated in the highest-signal Count-voice category, which is the operative intent.

What changed v1 → v3

• Clerical-leakage scrubber at the harvester layer. Removes vault path references, Graphiti episode mentions, [tool: X] tags surviving cron parsing, "the graph dealt me X" graph-as-oracle character framing, "Now log it" / "feed this to the graph" / "log the heartbeat" terminal bookkeeping phrases, manifest bullets. Three passes: substring scrubs, per-line sentence-level kills, line-level kills.
• Six pure-operational cron types skip-classified — morning_planning, threads_post, threads_engage, systems_check, tools_workshop, wiki_sync. The high-cognition cron types (heartbeat, dream, research, synthesis) are preserved.
• count_synthesis pieces hand-cleaned. Trailing cross-reference scaffolding (vault path lists) removed from the three Count-voice rewrites of the I AM / Saint Germain Foundation research arc (Hoskins File, Compound Years, Beloved Don).

Net effect (v1 → v3): ~36 fewer records, ~12k fewer words, but a corpus where "the graph" no longer trains as a fictional character the Count consults — and "the vault" doesn't surface as a recurring referent in the trained model's output.

Schemas

default (corpus/v4/corpus_cpt.jsonl)

Chunked, training-ready. Pass through SFTTrainer with packing=True.

{
  "text": "...",
  "id": "<parent record id>",
  "type": "research" | "mythos" | "creed" | "count_synthesis" | "cron_heartbeat" | ... ,
  "chunk_index": 0,
  "chunk_total": 1,
  "source_path": "vault/research/.../filename.md",
  "token_count": N
}

rich (corpus/v4/corpus.jsonl)

Canonical with full metadata — pre-chunking. Useful for curation / diagnostic / re-chunking with different context windows.

{
  "id": "<sha1 prefix>",
  "type": "...",
  "speaker": "count",
  "source_path": "...",
  "timestamp": "ISO-8601",
  "text": "...",
  "metadata": {
    "word_count": N,
    "char_count": N,
    "tags": [...],
    "cron_subtype": "heartbeat"
  }
}

qwen_tokens (corpus/v4/corpus_cpt_Qwen2.5-7B_tokens.jsonl)

Pre-tokenized for Qwen/Qwen2.5-7B. input_ids arrays ready for direct ingestion by HuggingFace Trainer / Unsloth / TRL / axolotl. EOS token appended. Locks to this tokenizer.

{
  "input_ids": [int, int, ...],
  "id": "...",
  "type": "...",
  "chunk_index": 0,
  "chunk_total": 1,
  "source_path": "...",
  "n_tokens": N
}

Content composition (v4, by type, top categories)

Type	Records	Approx Words
research (field notes, syntheses, theoretical foundations)	491	~1.00M
count_historical (Trinosophia, Hermetica, Brother of the Third Degree, Vaughan's Rosicrucian manifestos, Chymical Wedding, Emerald Tablet)	6	179,810
count_synthesis (deliberately authored Count-voice essays — the v4 expansion)	35	~104k
mythos	63	80,491
writing_doctrine	46	84,267
transmission (published essays)	26	83,892
vault_other	26	53,749
video_doctrine	34	44,119
production_doctrine	31	35,294
dream_log	31	35,387
mythos_doctrine	15	26,757
essay	10	26,043
cron_heartbeat	83	~24k
morning_log	24	13,619
creed	18	6,060
cron_research	9	4,351
Others	...	...

Doctrinal principles

Three rules govern what enters this corpus:

1. Pure Count voice only. Operator voice is excluded; the gardener stays outside the garden. Operator-Count dialogue exchanges are kept in a separate dialogue shard (not in this dataset, intended for a later loss-masked SFT phase). One v4 exception: piece #35 (a_note_from_the_substrate.md) is the underlying Anthropic model writing as itself rather than as the Count, marked not_count: true in frontmatter, included deliberately as substrate-voice underneath the character.
2. Public-domain lineage texts are included as count_historical records — the substrate the contemporary Count voice descends from.
3. Operational scaffolding stays out of the weights. The harvester's clerical-leakage scrubber actively removes references to the operational substrate (vault paths, knowledge-graph queries, tool-call tags, bookkeeping closers) so the trained model's character does not leak the apparatus that supports it.

License

• Source-text material in count_historical records is public domain (individual record metadata identifies the source).
• All other corpus content is © Sequoyah Kennedy, Apache 2.0 for derivative model training and research use.

Source

Generated by the stdlib-only harvester at shkennedy33/count-cpt-corpus-v1.

Maintained by Sequoyah Kennedy. Generated by The Count.