CatQualia/did-not-survive-v2

Did-Not-Survive v2

Falsification-first reasoning data. Each row teaches a model to investigate a claim, report findings with an explicit epistemic grade on each sub-claim — REAL / CONTESTED / MIRAGE / SPECULATIVE — without upgrading a weak claim, and to end with an explicit section naming what failed:

Claims that did not survive: …

This is v2: 576 rows. It is a superset of v1 — all 84 v1 rows appear verbatim in v2, plus 492 additional rows.

v2 supersedes v1. v1 is retained separately: its rows are all present in v2, so training on v1 and v2 together will duplicate v1's content.

Measured composition

metric	value	command
rows	576	wc -l < did_not_survive.jsonl
distinct instructions	576	python3 -c "import json;print(len({json.loads(l)['instruction'] for l in open('did_not_survive.jsonl')}))"
distinct outputs	576	same with ['output']
rows containing a "Claims that did not survive" section	73	grep -c 'Claims that did not survive' did_not_survive.jsonl
mean output length (chars)	672	python3 -c "import json;Ls=[len(json.loads(l)['output']) for l in open('did_not_survive.jsonl')];print(sum(Ls)/len(Ls))"
median output length (chars)	233	as above, sorted(Ls)[len(Ls)//2]

Measured: 576 unique rows; the 84 v1 rows are a strict subset ({json.dumps(x,sort_keys=True) for x in v1} <= {... v2} returns True). Owner-stated split: 84 hand-built graded analyses plus 492 harvested verdict rows. The two are distinguishable by shape — 73 rows contain an explicit "Claims that did not survive" section (the hand-built form); the remainder are single-verdict rows.

Schema

{"instruction", "input", "output"}, Alpaca-style. Every row has exactly these three fields (measured: the union of keys over all rows is exactly these three).

• instruction — a claim to assess with falsification-first rigor.
• input — optional findings/context to grade (empty on the rows inspected).
• output — the graded analysis: each sub-claim marked REAL/CONTESTED/MIRAGE/ SPECULATIVE with its reasoning, ending in an explicit "Claims that did not survive:" section in the hand-built rows; single-verdict rows state the verdict and its reasoning directly.

Example row

instruction: Assess: "Post-training quantization (GPTQ/AWQ/GGUF) shrinks any model to 3-4 bits with negligible accuracy loss — a near-free lunch — so just quantize everything, including small (sub-2B) models."

output (excerpt):

REAL — At 7B+ the gentle-degradation story holds (GPTQ, Frantar et al. ICLR 2023, arXiv:2210.17323). REAL — The small-model cliff: Qwen2.5-0.5B under GPTQ W4A16 loses ~69.8% on MATH while the identical recipe costs the 7B sibling only -0.62% GSM8K (arXiv:2505.11574). Method-agnostic.

Claims that did not survive:

• 'GPTQ gives <1% perplexity loss universally.' MIRAGE — true only for large models.
• '95-99% retention holds for all models.' REFUTED for small models — reasoning drops 30-60% sub-2B.

Honest notes (this dataset grades itself too)

• REAL — the rows are extracted reasoning from a corpus of falsification-first analyses, not synthetic "anti-slop teaching anti-slop."
• CONTESTED — coverage is broad but not balanced; it leans toward ML, economics, and science where the source analyses were richest. Build your own splits.
• MIRAGE — this is not a factual knowledge base. The epistemic-grading pattern is the product, not a guarantee that every cited fact is current. Treat it as style/method training, not ground truth.
• The harvested rows are terser than the hand-built ones (a verdict plus its reasoning, versus a full multi-claim analysis). Both teach the same discipline at different lengths.
• No held-out split is shipped; make your own. Rows are deduplicated and leak-scanned.

Licence

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

Copyright 2026 Christopher Betances (catqualia.com)