nnilayy/pg19-stability-bench

PG-19 Stability Evaluation Prompts

Long-context prompts in chat-message format at five bucket sizes (8K, 16K, 32K, 64K, 128K user-message tokens), designed for output-stability evaluation of LLMs under stress: as context grows (and RoPE scaling extends the effective window), do generations stay coherent — or do they degrade into mojibake, token soup, phrase loops, or script drift?

Each prompt is a 2-turn conversation (system + user) ready to send to any OpenAI-compatible /v1/chat/completions endpoint. The system message asks the model for a 500-word summary plus a brief outlook, producing ~600–800 generated tokens — enough output for mechanical degradation detectors (token soup, token flood, phrase loop, stuck token, script drift) to operate on.

Compared to nnilayy/pg19-concurrency-bench, which targets throughput measurement (5-word summaries, tightly bounded output, 256 prompts per bucket), this dataset targets output-quality at long context — longer generations, hand-picked novels long enough to fill 128K tokens, narrower prompt count.

Quick preview

system  : You are a helpful assistant. You will be given a passage from a
          book. Read it carefully, then summarize it in exactly 500 words
          and give a brief outlook on it.

user    : Context:

          [N tokens of cleaned PG-19 prose — N varies per config]

          Question: Summarize the above passage in exactly 500 words and
          give a brief outlook on it.

What the dataset viewer renders

Tab	What you see
Config dropdown (top)	Switch between 008k / 016k / 032k / 064k / 128k (default 008k)
Rows table	Per-prompt rows — the messages column auto-renders as chat bubbles
Statistics (per column)	Histograms over tokens and chars; length stats over text fields
SQL Console	Query in browser via DuckDB, e.g. SELECT title, tokens FROM train ORDER BY tokens DESC

Source

Sourced from emozilla/pg19 — a Parquet mirror of DeepMind's PG-19, containing books from Project Gutenberg published before 1919.

Construction

Ten long English-prose novels (published 1880–1915, all > 850K chars raw, mixed UK/US authors), hand-picked for variety and to ensure every book is long enough to fill a 128K-token bucket with margin:

gid	title
2145	Ben-Hur: A Tale of the Christ by Lew Wallace
21249	Clayhanger by Arnold Bennett
217	Sons and Lovers by D. H. Lawrence
35338	Marriage by H. G. Wells
31824	The Genius by Theodore Dreiser
10038	The Magnetic North by Elizabeth Robins
31858	Ancestors by Gertrude Atherton
10064	Beltane The Smith by Jeffery Farnol
31620	Vashti by Augusta J. Evans Wilson
27618	The End of a Coil by Susan Warner

Boilerplate cleaned by skipping front matter (table of contents, dedication, transcriber notes) to the first CHAPTER / PROLOGUE / BOOK ONE / PART ONE marker, so every passage starts in narrative prose, not on a title page.

Tokenized with tiktoken o200k_base (GPT-4o tokenizer).

For each bucket, a passage of (target_tokens − prefix − suffix) tokens is decoded back to text and wrapped as:

Context:

{passage}

Question: Summarize the above passage in exactly 500 words and give a brief outlook on it.

That user message is paired with a fixed system prompt to form a 2-message conversation.

The same 10 stories appear in every bucket, sliced to different lengths — clean apples-to-apples comparison across context sizes (only the prompt length varies between buckets).

Schema

Each JSONL record:

{
  "id": "pg19_000",
  "title": "Ben-Hur: A Tale of the Christ by Lew Wallace",
  "source_url": "http://www.gutenberg.org/ebooks/2145",
  "tokens": 8192,
  "chars": 30924,
  "messages": [
    {"role": "system", "content": "..."},
    {"role": "user",   "content": "Context:\n\n[passage]\n\nQuestion: ..."}
  ]
}

tokens counts the user message only (system message + chat-template wrappers are fixed overhead per model and don't vary across buckets).

Configs

Config	Target tokens	# prompts
008k	8,192	10
016k	16,384	10
032k	32,768	10
064k	65,536	10
128k	131,072	10

Usage

from datasets import load_dataset
from openai import OpenAI

ds = load_dataset("nnilayy/pg19-stability-bench", "032k", split="train")
client = OpenAI(base_url="http://localhost:8000/v1", api_key="dummy")

resp = client.chat.completions.create(
    model="your-model",
    messages=ds[0]["messages"],
    max_tokens=1024,
    temperature=0.0,
)
print(resp.choices[0].message.content)

Intended use

Designed for output-stability evaluation at long context, not for general accuracy benchmarking. The "500-word summary + brief outlook" task is a forcing function that produces substantial generation (~600–800 tokens) so mechanical degradation detectors — token soup (mojibake / non-ASCII), token flood (single-char repetition), phrase loop (multi-token n-gram repetition), stuck token (function-word collapse / low lexical diversity), and script drift (writing-system switch) — have enough output to operate on.

Pairs naturally with RoPE-extended models (e.g. Qwen with YARN) where the 128K bucket exercises the extrapolated context window beyond the model's native train-time length.

License

Apache 2.0 (matching the upstream PG-19 license).