Datasets:

DimSimd
/

kairos-thinking-5k

language:
  - en
license: cc0-1.0
task_categories:
  - text-generation
pretty_name: KAIROS-Thinking-Abduction-5k
size_categories:
  - n<10K
tags:
  - abductive-reasoning
  - chain-of-thought
  - scientific-reasoning
  - cross-domain-analogy
  - synthetic
  - instruction-tuning
configs:
  - config_name: default
    data_files: data/eureka_kairos_v1_5k.jsonl

KAIROS-Thinking-Abduction-5k

5,001 synthetic examples that teach abductive reasoning: how to pivot from a failed hypothesis to a cross‑domain breakthrough.

Most instruction datasets reward deductive step‑by‑step answers. KAIROS targets the other half of scientific thinking — the Eureka moment: recognising that the current frame is wrong, borrowing an analogy from an unrelated field, and proposing a novel but testable hypothesis.

This is a mirror of the public Kaggle dataset dimsimd/kairos-thinking-5k (CC0‑1.0), with two malformed JSON lines repaired and a dataset card added.

Dataset structure

One JSON object per line, 13 fields, identical across all records.

Field	Type	Description
`discovery_topic`	`string`	The scientific breakthrough or invention being reasoned about
`category`	`string`	Domain label (e.g. `Physics`, `Computer Science`, `Space Technology`)
`context`	`string`	Technical background, historical setting and constraints
`standard_logic_failure`	`string`	Why conventional deductive reasoning reached a dead end — the mental block
`failed_hypothesis`	`string`	The specific approach that was tried and failed
`why_it_failed`	`string`	Root‑cause analysis of the failure
`correction`	`string`	The conceptual shift required to escape the dead end
`the_pivot`	`string`	The crucial insight — usually a cross‑domain analogy
`reasoning_trace`	`string`	Chain‑of‑thought style monologue: confusion → clue → re‑framing → clarity
`inventive_hypothesis`	`string`	The resulting breakthrough solution
`novelty_score`	`float`	Self‑reported novelty, 0.6 – 0.95 (mean 0.81)
`cross_domain_distance`	`list[string]`	1–3 fields the analogy was drawn from (mean 3.0)
`implementation_feasibility`	`float`	Self‑reported feasibility, 0.2 – 0.9 (mean 0.71)

Statistics


Records	5,001
Unique `discovery_topic` values	4,434
Category labels	36 raw variants (informal taxonomy, some near‑duplicates: `Space Technology` / `Space Tech`, `Medicine & Biotechnology` / `Medicine`)
Total text	~4.04 M words
Average record	~807 words · ~6.2 KB of JSON
Language	English
File size	31.0 MB (`data/eureka_kairos_v1_5k.jsonl`)

Largest categories: Computer Science (930), Engineering (809), Physics (637), Chemistry (608), Space Technology (584), Medicine & Biotechnology (418), Mathematics & Cryptography (283), Materials Science (173), Nanotechnology (87), Quantum Computing (75).

Example

First record, with long strings truncated (marked …):

{
  "discovery_topic": "Transistor",
  "category": "Physics",
  "context": "In the early 20th century, the primary challenge in electronics was to find a reliable alternative to vacuum tubes, which were bulky, consumed high power (typically 5-10 watts per tube), and had limited lifespans often less than 1000 hours. …",
  "standard_logic_failure": "The fundamental flaw in deductive reasoning was the assumption that vacuum tubes could be indefinitely improved to meet the demands of modern electronics. …",
  "failed_hypothesis": "Attempted to improve vacuum tube performance using barium-strontium oxide-coated cathodes and improved grid structures …",
  "why_it_failed": "The improved vacuum tubes still suffered from high power consumption …",
  "correction": "The fundamental shift involved moving from vacuum-based electronics to solid-state devices …",
  "the_pivot": "The cross-domain analogy came from observing the behavior of semiconductors in radio receivers and the study of quantum mechanics. …",
  "reasoning_trace": "Initial observations of semiconductor materials showed that they could rectify electrical signals, but their behavior was inconsistent and poorly understood. …",
  "inventive_hypothesis": "The breakthrough idea was to use a semiconductor material with precisely controlled impurities (doping) to create a three-layer structure (p-n-p or n-p-n) that could amplify electrical signals. …",
  "novelty_score": 0.9,
  "cross_domain_distance": ["solid-state physics", "quantum mechanics", "materials science"],
  "implementation_feasibility": 0.6
}

Usage

from datasets import load_dataset

ds = load_dataset("DimSimd/kairos-thinking-5k", split="train")
print(ds[0]["the_pivot"])

Ready‑made chat / SFT formatting (the teacher prompt suggested by the dataset author):

SYSTEM_PROMPT = (
    "You are Project KAIROS. When faced with a problem, use Abductive "
    "Reasoning and Cross-Domain Analogies inside your thinking process."
)

def to_chat(row):
    assistant = (
        f"## Context\n{row['context']}\n\n"
        f"## Why standard logic fails\n{row['standard_logic_failure']}\n\n"
        f"## Failed hypothesis\n{row['failed_hypothesis']}\n"
        f"### Why it failed\n{row['why_it_failed']}\n\n"
        f"## Correction\n{row['correction']}\n\n"
        f"## The pivot\n{row['the_pivot']}\n\n"
        f"## Breakthrough hypothesis\n{row['inventive_hypothesis']}"
    )
    return [
        {"role": "system", "content": SYSTEM_PROMPT},
        {"role": "user", "content": f"Challenge: {row['discovery_topic']}.\n{row['context']}"},
        {"role": "assistant", "content": assistant},
    ]

The raw reasoning_trace field can also be wrapped in <think>...</think> blocks if you are training an explicit reasoning model.

Intended use

SFT / LoRA fine‑tuning for scientific and engineering reasoning
Training "failure → re‑framing → hypothesis" behaviour rather than answer recall
Benchmarks and prompts exploring cross‑domain analogy generation
Research on synthetic reasoning data and self‑correction

Limitations

Fully synthetic, generated by Mistral‑Medium from expert‑system prompts. Historical narratives are plausible but not fact‑checked; treat context and failed_hypothesis as narrative scaffolding, not as a historical source.
novelty_score and implementation_feasibility are model self‑assessments on a narrow range, not calibrated measures.
Category labels are inconsistent (36 raw variants, including one‑off slash‑combined labels).
Duplicate themes exist across discovery_topic values; 4,434 unique topics over 5,001 records.
English only.

Provenance and repairs

Source: https://www.kaggle.com/datasets/dimsimd/kairos-thinking-5k, original file eureka_kairos_v1_5k.jsonl (31,039,929 bytes, 5,000 physical lines).

The upstream file contains two corrupt lines. Both were repaired without dropping any content, giving 5,001 valid records:

Line 4401 — duplicated object prefix: {"disc{"discovery_topic": "Biohybrid Robots", .... Removed the extra {"disc.
Line 5000 — two records concatenated, the second missing its opening {"discov: ...0.7}ery_topic": "Cold Storage via DNA Data Archiving", .... Split into two objects and restored the truncated key, yielding one extra valid record.

The untouched upstream file is kept at original/eureka_kairos_v1_5k.jsonl for verification. No other bytes were changed; field values, ordering and text are as published.

License

CC0‑1.0 — same as the upstream Kaggle release.

Citation

@dataset{kairos_thinking_5k_2026,
  title  = {KAIROS-Thinking-Abduction-5k},
  author = {DimSimd},
  year   = {2026},
  url    = {https://huggingface.co/datasets/DimSimd/kairos-thinking-5k},
  note   = {Mirror of https://www.kaggle.com/datasets/dimsimd/kairos-thinking-5k}
}