Instruction data for mechanism transfer: given a problem in one field plus a named mechanism from a different field, describe the borrowed mechanism and then turn it into a concrete application in the target field.

Measured composition

metric	value	command
rows	12	`wc -l < cross_domain_isomorphisms.jsonl`
distinct instructions	12	`python3 -c "import json;print(len({json.loads(l)['instruction'] for l in open('cross_domain_isomorphisms.jsonl')}))"`
distinct `(domain_a, domain_b)` pairs	12	`python3 -c "import json;print(len({(json.loads(l)['domain_a'],json.loads(l)['domain_b']) for l in open('cross_domain_isomorphisms.jsonl')}))"`
distinct `mechanism` values	12	same with `['mechanism']`
rows with empty `input`	12	as above on `input`

Each of the 12 rows carries a distinct domain pair and a distinct mechanism — there is no repetition of either within the file. With 12 rows and 12 distinct pairs, every row is a separate transfer instance.

Schema

One JSON object per line:

field	rows containing it
`instruction`	12
`input`	12
`output`	12
`domain_a`	12
`domain_b`	12
`mechanism`	12

instruction — the target problem, the mechanism to borrow (usually with its source field named in brackets), and the instruction to describe the mechanism before applying it.
domain_a / domain_b — the two fields the transfer spans.
mechanism — the borrowed mechanism, named.
input — auxiliary context (empty throughout).
output — the borrowed mechanism, then the concrete application.

Example row

instruction: You manage a directed supply network (material flows supplier -> buyer). You want to predict how far a disruption cascades, but the standard robustness scalars from network science (algebraic connectivity, effective-graph-resistance / 'network criticality') keep mispredicting which failures blow up. Is there a mechanism from another field that predicts cascade extent better, and …

Domain pairs present

domain_a	domain_b
`abstract operational shapes of fictional/narrative mechanisms`	`computational drug repurposing / target identification`
`adaptive immunology (thymic negative selection)`	`one-class anomaly / novelty detection`
`complex/Hermitian spectral operators (physics: charged particle in a magnetic field)`	`finding communities in a directed network where the groups are defined by flow, not density`
`directed spectral geometry / phase-coherence operators`	`antibody / T-cell repertoire clonal-family recovery`
`directed spectral graph theory / food-web ecology`	`supply-chain resilience`
`directed/asymmetry-aware spectral operators`	`neural-network pruning at high sparsity`
`extremophile biology (cryptobiosis / anhydrobiosis)`	`compressing a large stateful system to a small revivable snapshot`
`olfactory neuroscience / chemistry of smell`	`building similarity embeddings where distance must match human judgement`
`plasma physics / MHD propulsion`	`building a self-clocking pump or valve with no external timing circuit`
`population genetics / ecology`	`training models on their own (synthetic) outputs`
`spectral graph partitioning`	`low-bit weight quantization difficulty`
`statistical physics of correlated systems (direct coupling analysis)`	`predicting which residues in a protein are in physical contact`

Honest notes

Small and curated. 12 rows. This is a seed set for the transfer pattern, not a bulk corpus.
Provenance of the answers: the outputs are model-written reasoning accompanying each transfer; they are not independently verified for factual correctness. Treat as style/method training for the transfer pattern, not as ground truth about any of the cited fields.
The instruction text names real, published mechanisms. Whether each borrowed mechanism is genuinely applicable to its target domain has not been independently evaluated — no row carries a validity verdict.
No held-out split is shipped; make your own.
_raw.jsonl is byte-identical to cross_domain_isomorphisms.jsonl (verified: both 33080 B, md5sum identical). It is the pre-rename artefact retained for provenance. Load only one of the two — using both double-counts every row.

Files

cross_domain_isomorphisms.jsonl — 33080 B, 12 rows.
_raw.jsonl — 33080 B, byte-identical duplicate of the above.
_build_rows.py — row builder.
_leak_scan.py — regex leak scanner.
LICENSE — CC-BY-4.0.

Licence

CC-BY-4.0. Free to use with attribution, including commercially.

Downloads last month: 9

Total file size:

126 kB