ksopyla/cogito-probe-bind

CogitoProbe-Bind: who-has-what entity binding

Synthetic people-and-attributes QA: each name gets a job, a city, a colour, and a friend. The model must answer who has which colour, who lives where, or where a person's friend lives. A bag-of-words embedding is not enough when everyone shares the same attribute vocabulary.

Author: Krzysztof Sopyła · License: Apache-2.0 · Seed: 20260916 · Tokenizer: HuggingFaceTB/SmolLM3-3B

In 60 seconds

Each entity is a bundle of attributes, written as:

alice is baker . alice in paris . alice color red . alice friend bob .
bob is miner . bob in oslo . bob color green . bob friend alice .

Then filler, then one of three questions (packed over several names):

task	Query looks like	Answer
attr_color	Q color alice bob	red green
who_place	Q who in paris oslo	alice bob
hop_friend_place	Q hop place alice	oslo (Bob's city)

If colour lookup works but the friend-hop stays at chance, the model stored labels, not bindings.

Load it

from datasets import load_dataset

ds = load_dataset("ksopyla/cogito-probe-bind")
row = ds["validation"][0]

print(row["seq_len"], row["task"], row["variant"])
print("query: ", row["query"])
print("answer:", row["answer"])
print("prize bits:", row["prize_bits"], "gap:", row["gap"])

# Loss only on the answer span (already marked).
# input_ids / labels are lists of int, length == seq_len.
loss_tokens = [t for t in row["labels"] if t != -100]

# Start small on a laptop: 1,024-token rows, fixed fact count.
small = ds.filter(lambda r: r["seq_len"] == 1024 and r["variant"] == "fixed")

You can ignore input_ids and train from context / query / answer as text. If you do use the provided ids, they are already tokenized for HuggingFaceTB/SmolLM3-3B (Llama-3 vocab) and must not be re-tokenized.

The four CogitoProbe datasets

Dataset	Job in one line	Typical use
ksopyla/cogito-probe-bits	Recall values for keys buried in a haystack	Memory / retrieval / compression capacity
ksopyla/cogito-probe-bind	Who has which colour, who lives where, friend's city	Compositional binding vs bag-of-words
ksopyla/cogito-probe-arith	Nested arithmetic + bracket matching	Did it store the expression tree?
ksopyla/cogito-probe-props	Object colours amid fluent filler	Facts vs padding statistics

Length ladder (every family): 1024 → 4096 → 8192 → 16384 → 32768. Half the rows are fixed (same fact count as at 1k, longer haystack), half are scaled (more facts as the row grows).

Why these exist

Web text is locally predictable: a language model can look strong by guessing nearby words without remembering a fact from thousands of tokens earlier. These four datasets hide a known set of facts in a long padded haystack so you can measure whether a model (or a small latent memory) actually stored them.

Each row tells you how many bits the answer is worth (prize_bits) and how far the question sits from the last fact (gap). That is the whole point: the information content is labelled, the distractor text is not the prize, and the length is a ladder rather than a single context size.

Real rows use random single-token English-ish pieces from the Llama-3 / SmolLM3 vocabulary (gonzalez, oslo, validators, …), not the toy names alice / bob in the examples above. The grammar of the task is the same.

How to score

Train or evaluate only on the answer span. Teacher-forced token accuracy on labels != -100 is the main number. Recovered bits against the labelled prize:

max(0, prize_bits + Σ log2 p(gold_t))

A decoder that cannot see tokens more than gap away must sit at chance — the evidence is that far from the answer.

Report attr_color, who_place, and hop_friend_place separately. Colour-only success with hop at chance means the model stored a list of colours, not who-has-what. A strong extra check: swap one entity's colour in meta and require that entity's answer token to flip.

Schema

column	meaning
text / context / query / answer	Readable surfaces. text is the full padded row.
input_ids, attention_mask, labels	Ready for causal LM training. labels is -100 everywhere except the answer.
seq_len, rung	Padded length: 1024, 4096, 8192, 16384, or 32768.
variant	fixed = same number of facts as at 1k, longer haystack. scaled = more facts as the row gets longer.
task	Question type inside this family (see above).
prize_bits	Known information content of the gold answer (combinatorial lower bound).
gap	Tokens from the last evidence token to the start of the answer.
answer_start / answer_end / evidence_end	Character-free token indices into input_ids.
meta	JSON string: fact table, fingerprints, node values.

This build

split	rows
train	8448
validation	896
test	896

metric	value
total rows	10240
token length (all padded)	min 1024 / p50 4096.0 / max 32768
mean prize bits	58.381 (min 24.000, max 160.000)
mean gzip ratio (text)	0.037 (n=160 stratified sample)
mean gzip ratio (int32 ids)	0.050
mean unigram entropy (bits)	6.279
mean bigram entropy (bits)	6.488
answer entropy (bits)	13.322 over 10240 strings
tasks	{'attr_color': 3494, 'who_place': 3491, 'hop_friend_place': 3255}
variants	{'scaled': 5120, 'fixed': 5120}
rungs	{'seq1024': 4608, 'seq4096': 2560, 'seq8192': 1280, 'seq16384': 1024, 'seq32768': 768}

Per-rung means:

seq_len	n	mean prize bits	mean gap	mean gzip(text)
1024	4608	34.538	887.0	0.090
4096	2560	54.600	3891.0	0.036
8192	1280	97.312	7851.0	0.031
16384	1024	97.188	16043.0	0.018
32768	768	97.417	32427.0	0.011

Example rows (truncated):

• bind/seq1024/scaled/train/00000 task=attr_color prize=40.00 bits gap=887 query=Q color idx wish oxidation neben graf morr spacer travellers answer=green coral mouse dog oslo bird amber yellow
• bind/seq1024/scaled/train/00001 task=attr_color prize=40.00 bits gap=887 query=Q color auswahl youngsters avan urging processing morr navy neben answer=deer dog bird yellow pink white brown silver
• bind/seq1024/scaled/train/00002 task=who_place prize=24.00 bits gap=887 query=Q who in map miner box judge rope door actor rider answer=bn keep youngsters manifest include hx illustrations kerr

Split leakage

pair	fingerprint overlap	input_ids overlap	text overlap	answer-string overlap
train∩validation	0	0	0	0
train∩test	0	0	0	0
validation∩test	0	0	0	0

Within-split duplicate input_ids counts: {'train': 0, 'validation': 0, 'test': 0}.

Train / validation / test use disjoint random streams. A fingerprint of the facts is checked for overlap. Shared answer strings (for example the same 8 colours) are expected and are not a leak.

Rebuild

Deterministic rebuild (does not upload):

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

Ids are composed from a verified 1-token atom table of HuggingFaceTB/SmolLM3-3B. Arithmetic rows inject bare digit and bracket ids; they do not BPE-encode a glued string such as (1+2)*[3-4] (that merge path is not a well-defined alphabet).

Limitations

• Not natural language. Atoms are verified 1-token pieces of the SmolLM3 / Llama-3 vocab, chosen so each symbol is one id. Do not treat this as a human corpus.
• Answers are packed (several values in one span). Single-token labels are too sparse for a small latent channel to learn from.
• prize_bits is a counting lower bound on the answer, not a cross-entropy floor of a local language-model window.
• Arithmetic mixed brackets colour the tree; they do not change + - * meaning. eval-only accuracy is not evidence of rich structure.
• Rows are padded with a repeating filler cycle, so gzip of the full text looks tiny. Compare prize_bits, not compressibility of the padded row.

Origin

These files were built for a research project on compressing long context into a small set of latent vectors (“concepts”), so the author could ask what those vectors actually store. You do not need that project, its training code, or its internal experiment log to use the datasets.

Project page: ai.ksopyla.com · author: Krzysztof Sopyła. Generator: data/concept_probes/ in the public research repo (MIT).

License

Apache-2.0 for this synthetic dataset. No web scrapes, no personal data. Generator code is MIT.

Citation

@misc{cogitoprobe2026,
  title  = {CogitoProbe: synthetic long-haystack probes for memory and compression},
  author = {Sopyła, Krzysztof},
  year   = {2026},
  url    = {https://huggingface.co/datasets/ksopyla/cogito-probe-bind},
  note   = {Seed 20260916. Four families: bits, bind, arith, props.},
}