baiango/bland-stories-labeled

Bland Stories Labeled

197,016 machine-generated short fiction stories, produced deliberately in a "bland" register (templated, stock-phrased, mode-collapsed prose) for use as a control/negative class in prose-quality work, with 3-vote LLM annotations for genre, world, energy, and tags.

Every row is one model draw that passed generation gates. Rows are not high-quality fiction by design — see Limitations.

At a glance

Rows	197,016
Usable for SFT (verdict in {style,pass})	196,997
Bytes	894,379,958
MD5	5c7e8e981d7c92dae5475c1ddb188379
n_tok total (Qwen3-0.6B-Base tokenizer, over text)	112,098,174 (mean 569/row)
Rows with derived prose field	101,228
License	CC0-1.0 (public domain)

Generation setup

• Generator: Qwen3-0.6B-Base GGUF Q5_K_S (mradermacher quant), llama.cpp.
• Sampler (pinned, identical for every draw): top_k 40, top_p 0.95, temperature 1.0, min_p 0.1. No repetition or n-gram penalties.
• Prompt (verbatim in the prompt field): Write a story in English. Genre: {parent}. Subgenre: '{genre}'.\n\n (the Subgenre clause is omitted when the leaf name equals the parent).
• Leaf space: each row draws from a (parent, genre) pair; the file contains 197,006 distinct pairs over 119,534 distinct parents.
• MODEL-STOPS rule: a draw ships only if finish_reason == "stop" (the model ended its own story) and the text ends with sentence punctuation (ends == true). All 196,997 usable rows satisfy both; 7 rows with finish_reason == "length" and 12 gate-failed rows are retained as verdict == "failed" for transparency.
• Retries: up to 60 fresh-seed draws per leaf pair; the winning attempt number is in draws.
• Reproducibility: not reproducible — seeds are recorded per row but the multi-host build does not support exact regeneration.

Schema

field	type	coverage	meaning
k	str	100%	per-row hex id
parent	str	100%	taxonomy parent used in the prompt
genre	str	100%	subgenre/leaf name used in the prompt
prompt	str	100%	exact generation prompt prefix
seed	int	100%	sampling seed of the shipped draw
draws	int | null	98.5%	attempt number of the shipped draw; null when absent
text	str	100%	verbatim model completion (the full raw text)
prose	str | null	51.4%	re-derived story body (style-strip pass); present iff verdict == "style"; always differs from text; null elsewhere
n_tok	int	100%	subword count of text under the Qwen3-0.6B-Base tokenizer (not of prose)
stop	str	100%	finish_reason: stop (197,009) or length (7, all verdict: failed)
ends	bool	100%	text ends with sentence punctuation
verdict	str	100%	pass (95,769: text accepted as-is), style (101,228: re-stripped into prose), failed (19: excluded)
gates	list[str]	100%	failed gate names on failed rows (e.g. ["loop"]); [] on all other rows
att	int	100%	attempts exhausted, real values (≥1) only on 9 failed rows; -1 sentinel elsewhere
review	list | null	21.9%	style flags that triggered the strip (e.g. ["style-heavy"]); null when absent
energy	str	100%	annotated energy
world	str	100%	annotated world
tags	list[str]	100%	0–5 tags, each Tag or Parent / Subtag
confidence	str	100%	label agreement: unanimous (179,379) or 2-1 (17,637)

File format (HF-ready)

• Every row carries all 19 schema keys in a fixed order (a uniform schema is required for HF's json→parquet conversion); absent values are null.
• Type-stable sentinels instead of schema-split columns: att == -1 = not applicable, gates == [] = no gates evaluated; prose, review, draws are plain null when absent.
• The 19 verdict == "failed" rows sit at the top of the file: they carry the only real att/gates values, and the Arrow JSON reader must see them in its first 10MB chunk to infer the right types. All other rows keep their original build order. Rows are identified by k, never by line number.

Annotations

3-vote labeling with inclusionai/ling-3.0-flash; a label ships as unanimous when all votes agree, else 2-1.

• energy: External 66%, Relational 34%
• world: Fantasy & Supernatural 56%, Contemporary 29%, Science Fiction 11%, Historical & Western 4%
• tags (share of rows containing the tag): Action & Adventure 66.3%, Mystery & Crime 48.8%, Drama 37.3%, Slice of Life 37.3%, Horror 32.4%, Romance 13.5%

Text statistics

Perplexity of the story text under Qwen3-0.6B-Base (400-row seeded sample, token-weighted): PPL 3.43, 95% CI [3.42, 3.49]. Reference points under the same scorer: news prose 18.5–20, hand-written repetitive drone 6.5, Austen 6.2, Melville 4.5, specific modern literary prose ~46.

• The corpus scores below Melville: it is formula, not loops — 0.0% of tokens sit inside any repeated word 8-gram (distinct-8-gram ratio 1.000), yet stock collocations dominate: "nestled between" appears in 21% of rows and the heroine name "Elara" in 14.5%.
• Stock-opener templates ("In the quiet town of X, nestled between…") are the intended target class, not a defect.

Intended use

• Negative/control class for prose-quality gates and blandness detection.
• Genre/world/energy-conditional SFT where the target register is explicitly generic fiction.

Limitations

• Mode-collapsed by design. Openers, names, and imagery follow small-model clichés; treat any quality judgment accordingly.
• Self-perplexity caveat. The PPL scorer above is the same model family as the generator, which biases PPL downward; the number characterizes template density, not absolute quality.
• prose exists on only 51.4% of rows; on the rest the verbatim text is the story body.
• Labels are 3-vote LLM annotations (91% unanimous), not human-verified.
• The 19 failed rows are included for transparency; exclude via verdict != "failed" for training.

Provenance

Produced and labeled by the baiango/blandv2-gen pipeline (private repo; single-code-version build, audit passed): generation drivers, style-strip, gates, and the taxonomy leaves (leaves.jsonl) live there, as does the labeling chain under labeling/ (classify_prod_v3.py, merge_corpus_labels.py, merge_corpus_gate.py, row_order.py, retok_n_tok.py). The PPL measurement (ppl_prose.py) is in the local genre_classify workspace, not committed.

License

CC0-1.0. The text is entirely machine-generated; no human authorship is claimed.