Multi-lens and metaphor-to-buildable reasoning SFT data. 1056 instruction/output pairs in Alpaca schema, teaching two reasoning styles that standard SFT corpora do not cover: multi-lens superposition (hold several extreme verdicts in parallel instead of collapsing to one) and metaphor-to-buildable (turn a loose analogy into a concrete, implementable step).

Measured composition

metric	value	command
rows	1056	`wc -l < reasoning_superposition.jsonl`
distinct instructions	953	`python3 -c "import json;print(len({json.loads(l)['instruction'] for l in open('reasoning_superposition.jsonl')}))"`
distinct outputs	1056	same with `['output']`
rows with empty `input`	1056	`python3 -c "import json;print(sum(1 for l in open('reasoning_superposition.jsonl') if json.loads(l)['input']==''))"`
mean output length (chars)	378	`python3 -c "import json;Ls=[len(json.loads(l)['output']) for l in open('reasoning_superposition.jsonl')];print(sum(Ls)/len(Ls))"`
median output length (chars)	144	as above with `sorted(Ls)[len(Ls)//2]`
rows matching the multi-lens keyword filter in the Usage snippet below	163	see Usage

All 1056 outputs are distinct. 953 of 1056 instructions are distinct, so 103 rows share an instruction with another row.

Schema

{"instruction", "input", "output"}, Alpaca-style. input is empty on all 1056 rows — the task is self-contained in instruction.

Example row

instruction: Reason by MULTI-LENS SUPERPOSITION. Do not give one verdict on "Is this refactor safe to ship?". Run several extreme lenses in parallel and report the heatmap of agreement, conflict, and paradox.

output:

Do not collapse to one verdict. Run lenses in parallel:

T_PERF: 0.9 safe (latency improves).

T_CORRECTNESS: 0.4 safe (one branch is untested).

T_MAINTAINER: 0.2 safe (the author leaves Friday; nobody else reads this DSL).

T_SECURITY: PARADOX (it both narrows and widens the trust boundary depending on cache-poisoning).

HEATMAP (the answer): safe along PERF, contested along CORRECTNESS, blocked along MAINTAINER, undecidable along SECURITY. The verdict is the vector, not a word: ship only if you own all four columns.

Honest notes (read before training)

Synthetic. Outputs are model-synthesized distillation, not human-gold and not independently verified for factual correctness. Treat as style/format training data, not a knowledge source.
No held-out split is shipped; make your own. Outputs are 100% unique after dedup, but near-duplicate instructions may exist (953 distinct of 1056).
The "general structured reasoning" bucket is the largest; the two headline families (multi-lens, metaphor) are a minority of the set. Filter by pattern if you want only those.
Suitable for SFT on the reasoning format. Not suitable as a benchmark or a factual reference.
Leak-scanned. Rows containing internal paths, filenames, or private references were dropped before release. One residual reference remains, measured: exactly 1 of the 1056 rows contains the string .claude/agents/, inside an illustrative sentence about running new agent definitions against a library of host-harming patterns. It is a directory name used as an example, not an exposed path with a user or host in it (0 rows contain /home/). Disclosed rather than silently patched, because editing a row after the fact would change the shipped data; drop the row before upload if you prefer a card that says zero.

Usage

from datasets import load_dataset
ds = load_dataset("json", data_files="reasoning_superposition.jsonl")

Filter to just the multi-lens family:

ds = ds.filter(lambda r: "lenses in parallel" in r["output"].lower() or "heatmap" in r["output"].lower())

That filter selects 163 of 1056 rows.

Licence

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

Downloads last month: 12

Total file size:

755 kB