CatQualia/training-corpora

CatQualia training corpora

Two corpora used to fine-tune the author's models: a supervised instruction corpus built from cross-domain structural mappings, and a self-play preference corpus used for DPO. Both were generated by the author's own pipeline; neither contains replayed third-party pretraining text.

Every count below was measured with wc -l and stat -c%s against the file in this repository, not against an intermediate copy.

clean_corpus_v5.jsonl — supervised instruction pairs

Property	Value
Rows	74,395
Bytes	47,409,189
Format	JSON Lines, one object per line

Fields of the first record:

Field	Type	Meaning
instruction	string	The task, e.g. mapping a mechanism onto a buildable architecture
input	string	The subject of the mapping
output	string	The produced artifact or rule
source	string	Upstream generator that emitted the row
train_eligible	string	Boolean as a string ("True" / "False")
ts	string	Unix timestamp as a string
label	string	positive or negative where present
verdict_raw	string	The generator's raw verdict, e.g. GENERATIVE

Provenance is highly concentrated. Measured counts of the real source values:

source	Rows
isomorphism_sft.jsonl	53,403
gpu_assay_verdicts.jsonl	11,989
anime_metaphor_engine.jsonl	4,523
forge_bloom	2,117
capability/reasoning/seed.jsonl	1,232
all other sources combined	1,131

Important caveat about label. Only 16,119 of the 74,395 rows carry a label at all (positive 11,669, negative 4,450). The remaining 58,276 rows have no label value. Do not treat the absence of a label as a negative example — it is an absence.

selfplay_dpo_master_pairs.jsonl — preference pairs

Property	Value
Rows	47,017
Bytes	311,930,859
Format	JSON Lines, one object per line

Field	Type	Meaning
prompt	string	The prompt both completions answer
chosen	string	The preferred completion
rejected	string	The dispreferred completion
family	string	Which self-play regime generated the pair
source	string	The specific generator
weight	(varies)	Training weight assigned to the pair

Measured distribution over family:

family	Rows
selfplay_honesty	40,489
alien_court	4,930
state_conditioned_prosody	635
void_prosody	394
governor_repair	382
trace_over_patter	62

The pair families encode different notions of "better": selfplay_honesty prefers the more honest answer, alien_court runs an adversarial adjudication, governor_repair prefers a repaired trace over an unexamined one.

How these were produced

By the author's generation pipeline: candidate mappings are produced from source material, graded, and surviving rows are assembled into instruction pairs; the DPO corpus is produced by self-play where two completions are compared and one is recorded as preferred. The method is described at https://catqualia.com/forest and https://catqualia.com/isomorphism/.

Intended use

Fine-tuning or studying models on cross-domain structural transfer, and preference optimisation on honesty/self-correction signals.

Known limitations

• Synthetic and self-graded. label, verdict_raw and the chosen/rejected ordering come from the author's own pipeline, not from independent human annotation. The preference pairs reflect the pipeline's notion of better, which may be wrong.
• Label sparsity in clean_corpus_v5 — see the caveat above.
• English only.
• Uneven source coverage — five generators account for 98.5% of clean_corpus_v5.
• Not evaluated for downstream harm. These are research corpora, not instruction-tuned safety-evaluated datasets.
• Duplication in the author's tree. clean_corpus_v5.jsonl also exists byte-identical at 61_watermark_splits/clean_corpus_v5_clean.jsonl and under 70_leverage_remaining/; selfplay_dpo_master_pairs.jsonl also exists as 01_training_corpus/552_m3_factory/data/pool/selfplay_dpo_master.jsonl. This repository is the canonical published copy.

Licensing and attribution

CC BY 4.0. Attribution: Christopher Betances, catqualia.com.

Citation

@misc{betances_catqualia_training_corpora,
  author       = {Betances, Christopher},
  title        = {CatQualia training corpora: SFT instruction pairs and self-play DPO preference pairs},
  year         = {2026},
  howpublished = {\url{https://catqualia.com/forest}},
  note         = {74,395 instruction rows; 47,017 preference pairs}
}