adugeen/personal-facts-msc

Personal Facts (MSC) — Multi-Dimensional Annotation

A manually annotated dataset of 2,779 personal facts sampled from the Multi-Session Chat (MSC) corpus, labeled across seven dimensions that jointly characterize a fact's topic, temporal anchoring, referent, lifetime, validity, and dialogue-continuation potential.

The scheme extends PeaCoK with two new top-level categories (Demographics, Possessions) and three new dimensions (Duration, Validity / Invalidity Reason, Followup), and operates on individual facts rather than fact pairs.

Splits

Split	Examples
train	2,223
test	556

The split is stratified by main_category (seed 42, 80/20). The 556-fact test set is the canonical gold-annotated evaluation set used in the accompanying paper.

Schema

Each row is a single utterance candidate fact, together with the dialogue turn preceding it (context) for disambiguation.

Field	Type	Description
text	string	The candidate personal fact.
context	string	Preceding utterance from the dialogue (1 turn).
main_category	string	Topical category. One of Preferences, Experience, Goals and plans, Routine activities, Possessions, Characteristics, Relationships, Demographics, None.
time	string	Temporal anchoring: Past, Present, Future, None.
referent	string	Whom the fact is about: Self, Other, None.
duration	string	Expected lifetime: Short-term, Long-term, None.
broken	string	Validity flag: No (valid) or Yes (invalid).
broken_reason	string	Failure mode when invalid: Opinion, Multiple facts, No fact, Not about self/known people (third person), Context Insufficient, None.
followup	string	Suitability for dialogue continuation: Yes, Maybe, None.
labels	list[str]	Compact one-letter encoding of the seven dimensions, e.g. ['main_category_P', 'time_P', 'referent_S', 'duration_L', 'broken_N', 'broken_reason_N', 'followup_N']. Provided as a convenience for multi-label models.

None in the non-validity dimensions indicates the field is undefined for invalid facts (those with broken == 'Yes'); broken_reason == 'None' is used when the fact is valid.

Label inventory

main_category (9): Preferences, Experience, Goals and plans, Routine activities, Possessions, Characteristics, Relationships, Demographics, None. The first five are inherited from PeaCoK; Preferences is split out from PeaCoK's characteristics; Demographics and Possessions are introduced by this scheme.

time (4): Past, Present, Future, None.

referent (3): Self, Other, None. Other covers facts about people personally known to the speaker (family, pets, friends).

duration (3): Short-term, Long-term, None. Distinguishes ephemeral facts (e.g. "I work this weekend") from enduring ones (e.g. "I am a diver").

broken / broken_reason (Validity / Invalidity Reason): a fact is invalid if it cannot stand alone as a personal-fact memory entry. Reasons:

• No fact — utterance contains no personal information.
• Opinion — generic opinion or statement, not a personal fact.
• Multiple facts — multiple facts bundled in one utterance; should be decomposed before storage.
• Context Insufficient — meaning depends on missing context (e.g. "I gave her flowers").
• Not about self/known people — third-person reference that cannot be attributed (e.g. "She likes Indian food").

followup (3): Yes, Maybe, None. Whether the fact opens a natural dialogue continuation hook.

Construction

• Source. All facts are drawn from the Multi-Session Chat (MSC) corpus.
• Sampling. Personal facts associated with dialogue responses are deduplicated by exact text match, embedded with BGE-M3 (dense, L2-normalized), and clustered into 1,000 K-Means groups. Up to 3 facts are sampled per cluster, yielding 2,779 candidates with broad topical coverage.
• Annotation. Each fact was annotated by the author in Label Studio following written guidelines. LLM pre-annotations (Gemma-3-27B, Qwen-3-VL-32B, GPT-5-mini) were used as anchors but every label was reviewed and corrected manually.
• Inter-annotator agreement. On a held-out 300-fact pilot, Fleiss' κ ranges from 0.749 for main_category to 0.458 for broken_reason; surface agreement on validity is 92.1%.

Class distributions

Train (2,223) / Test (556):

Dimension	Distribution (train → test)
main_category	Preferences 458 → 115 · Experience 308 → 77 · Routine activities 298 → 75 · Goals and plans 268 → 67 · Characteristics 230 → 58 · Demographics 144 → 36 · Possessions 143 → 35 · Relationships 35 → 8 · None 339 → 85
time	Present 1332 → 331 · Past 290 → 73 · Future 261 → 67 · None 340 → 85
referent	Self 1728 → 437 · Other 155 → 34 · None 340 → 85
duration	Long-term 1505 → 379 · Short-term 378 → 92 · None 340 → 85
broken	No 1884 → 471 · Yes 339 → 85
followup	None 1945 → 484 · Yes 165 → 37 · Maybe 113 → 35

Usage

from datasets import load_dataset

ds = load_dataset("adugeen/personal-facts-msc")
print(ds)
# DatasetDict({
#   train: Dataset({features: [...], num_rows: 2223})
#   test:  Dataset({features: [...], num_rows: 556})
# })

ex = ds["train"][0]
print(ex["text"], "→", ex["main_category"], ex["time"], ex["referent"])

Intended uses

• Training and benchmarking multi-head / multi-label classifiers for persona-fact characterization.
• Quality filtering of persona corpora (the broken / broken_reason fields directly support invalid-fact detection).
• Memory-policy research for dialogue agents: duration and followup are designed to drive long-term-memory eviction and topic-continuation policies.
• Auditing the label distribution of larger persona corpora — applying a classifier trained on this data to PersonaChat and full MSC reveals that ~30% of MSC facts are predicted invalid.

Limitations and biases

• Single annotator. The full corpus was annotated by one person; inter-annotator-agreement statistics come from a 300-fact pilot.
• Class imbalance. Relationships (≈1.5%) and the Yes followup class (≈7%) are heavily under-represented; metric reporting should use macro F1 rather than accuracy.
• Source bias. All facts come from MSC, which is itself crowd-authored English dialogue. Distributions and category boundaries may not transfer to other languages or domains (clinical, technical, etc.).
• broken_reason is the noisiest dimension. Annotator agreement is lowest here (κ = 0.458); downstream models trained on this dimension inherit that uncertainty.

License

Released under CC BY-SA 4.0, matching the source corpus (Multi-Session Chat). When using or redistributing, please also cite the underlying MSC dataset.

Citation

If you use this dataset, please cite the accompanying paper (citation will be added once the proceedings are published):

@misc{personal-facts-msc-2026,
  title  = {An Extended Annotation Scheme for Personal-Fact Classification in Dialogue},
  author = {Zaitsev, Konstantin},
  year   = {2026},
  note   = {Dataset: https://huggingface.co/datasets/adugeen/personal-facts-msc}
}

Please also cite the source corpus:

@inproceedings{xu-etal-2022-beyond,
  title     = {Beyond Goldfish Memory: Long-Term Open-Domain Conversation},
  author    = {Xu, Jing and Szlam, Arthur and Weston, Jason},
  booktitle = {Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics},
  year      = {2022}
}

Contact

Questions and issues: open a discussion on the dataset page or contact adugeen.