ksopyla/cogito-probe-bits

CogitoProbe-Bits: key→value recall in a long haystack

Part of Cogito Capability Checks — does your model actually use what it read? Small synthetic tests with exact answers, guessing floors and a length ladder (1k → 128k tokens). Project page: ai.ksopyla.com/projects/cogito-capability-datasets

TL;DR. key X val Y facts at the start, 1k–32k tokens of filler, then Q k1 k2 … asks for several values at once. Measures how many bits of facts a model (or a compressed memory) actually kept, and from how far away.

Lengths	1k, 4k, 8k, 16k, 32k tokens (seq_len)
Size	10,240 rows: train 8,448 · validation 896 · test 896
Tests	multi-key associative recall; memory capacity
Answer	the asked values, in order (32 possible values each)
Guessing	1/32 per value
Format	text + ready ids for HuggingFaceTB/SmolLM3-3B (Llama-3 vocab)
License	Apache-2.0

Example

key alice val red . key bob val blue . key carol val green .
… 1k–32k tokens of filler …
Q alice carol        →   red green

Quick start

from datasets import load_dataset

ds = load_dataset("ksopyla/cogito-probe-bits")            # train / validation / test
row = ds["validation"][0]
print(row["seq_len"], row["task"], "|", row["query"], "->", row["answer"])

# Already tokenized for HuggingFaceTB/SmolLM3-3B (Llama-3 vocab); loss only where labels != -100.
# Or ignore the ids and use the text fields: context / query / answer.
small = ds["train"].filter(lambda r: r["seq_len"] == 1024 and r["variant"] == "fixed")  # laptop-sized start

How to score

Teacher-forced token accuracy on the answer span (labels != -100) is the main number. Also report recovered bits = max(0, prize_bits + Σ log2 p(gold token)), plotted against prize_bits. fixed rows keep ~16 facts at every length; scaled rows grow to 256 facts at 32k. Sanity check before reading a memory result: a dense Transformer should exceed ~75 % at seq_len=1024.

The family: Cogito Capability Checks

One goal: check whether a model stores and uses facts from its context, not just predicts nearby words. Each dataset plants facts with known answers, adds filler, asks a question at the end, and labels how hard guessing is. Length is a ladder, so you see where a model stops working.

dataset	input	the question it answers	lengths
cogito-probe-bits	symbol keys and values	How many key→value facts did it keep? (memory capacity in bits)	1k–32k
cogito-probe-bind	people with attributes	Does it keep who has what together, and follow one link?	1k–32k
cogito-probe-arith	nested bracket arithmetic	Did it store the expression tree, not only the result?	1k–32k
cogito-probe-props	short facts in fluent filler	Does it keep the facts and ignore the padding?	1k–32k
cogito-text-world	simple English stories	Can a small LM trained from scratch quote, look up, track, chain, count and deduce facts in a story?	1k–128k

Start with a probe if you test a memory or compression module on a pretrained tokenizer; start with cogito-text-world if you compare architectures trained from scratch on language.

Related work and inspirations

• Needle in a Haystack (Kamradt, 2023) and passkey retrieval (Mohtashami & Jaggi, 2023, arXiv:2305.16300) — one fact in long filler; here: many facts, labelled bits, decoys.
• RULER (Hsieh et al., 2024, arXiv:2404.06654) — synthetic long-context tasks (multi-key, variable tracking) for pretrained LMs; these sets add from-scratch training data and evidence-removed controls.
• BABILong (Kuratov et al., 2024, arXiv:2406.10149) and bAbI (Weston et al., 2015, arXiv:1502.05698) — facts hidden in long natural text.
• Zoology / MQAR (Arora et al., 2023, arXiv:2312.04927) — multi-query associative recall as an architecture test.
• Lost in the Middle (Liu et al., 2023, arXiv:2307.03172) — answers depend on where the evidence sits.
• Binding IDs (Feng & Steinhardt, 2023, arXiv:2310.17191), ListOps (Nangia & Bowman, 2018, arXiv:1804.06028), PrOntoQA (Saparov & He, 2022, arXiv:2210.01240) — binding, tree structure and rule chains.
• TinyStories (Eldan & Li, 2023, arXiv:2305.07759) and SimpleStories (Finke et al., 2025, arXiv:2504.09184) — the simple language under cogito-text-world.

Schema, build statistics, rebuild, limitations

Columns

column	meaning
text, context, query, answer	readable surfaces; text is the full padded row
input_ids, attention_mask, labels	SmolLM3-3B ids, length = seq_len; labels = -100 except on the answer
seq_len, rung	1024, 4096, 8192, 16384 or 32768
variant	fixed: same facts as at 1k, longer filler (a drop = length failure); scaled: more facts with length (a drop = capacity failure)
task	question type (see above)
prize_bits	information in the gold answer (counting lower bound)
gap	tokens from the last evidence token to the answer
answer_start, answer_end, evidence_end	token indices into input_ids
meta	JSON: fact table, fingerprints, node values

Build (seed 20260916, 2026-09-16): 10,240 rows = train 8,448 / validation 896 / test 896. Rows per length: 1k 4,608 · 4k 2,560 · 8k 1,280 · 16k 1,024 · 32k 768; half fixed, half scaled. Mean prize_bits 89 (40–640); task recall_packed on every row. Splits use disjoint random streams; fact fingerprints, input_ids, text and answer strings have zero overlap between train, validation and test (shared answer values, such as colours, are expected).

Rebuild (deterministic):

uv run python scripts/build_concept_probe_datasets.py --scale full --seed 20260916 \
  --tokenizer HuggingFaceTB/SmolLM3-3B --families bits --out_dir Cache/concept_probes/full

Generator: data/concept_probes/ in github.com/ksopyla/MrCogito.

Limitations

• Not natural language: symbols are single tokens of the SmolLM3 / Llama-3 vocabulary (real rows use pieces like gonzalez, oslo, validators; the examples above use toy names).
• Answers are packed (several values per answer), so a small memory gets a dense learning signal.
• prize_bits is a counting lower bound, not a cross-entropy floor.
• The filler is a repeating cycle, so gzip of the whole row is tiny; compare prize_bits, not compressibility.

Changelog

date	version	change
2026-10-10	card	Card rewritten: one-line goal, quick start, family table, related work, citations. Data unchanged.
2026-09-16	v0	First public release, seed 20260916, 10,240 rows.

Citation

Please cite the family, and the dataset you used:

@misc{sopyla2026cogitochecks,
  title        = {Cogito Capability Checks: synthetic tests of whether models store and use facts from long context},
  author       = {Sopyła, Krzysztof},
  year         = {2026},
  howpublished = {\url{https://ai.ksopyla.com/projects/cogito-capability-datasets/}},
  note         = {Datasets: cogito-probe-bits, -bind, -arith, -props; cogito-text-world}
}

@misc{cogitoprobe_bits,
  title        = {CogitoProbe-Bits: key-value recall in a long haystack},
  author       = {Sopyła, Krzysztof},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/ksopyla/cogito-probe-bits}},
  note         = {Cogito Capability Checks; seed 20260916, data v0}
}

Author: Krzysztof Sopyła · code: github.com/ksopyla/MrCogito.